(Analysis by Benjamin Qi, Sanjeev Murty)

Call the number of steps that a permutation requires the period of the permutation. Every permutation can be partitioned into cycles of sizes $c_1,c_2,c_3,\ldots,c_k$ such that $c_1+c_2+\ldots+c_k=N$. Then the period is equal to $\text{lcm}(c_1,c_2,\ldots,c_k)$.

Subtask $N\le 50$:

Maintain the number of possible permutations for each possible LCM of a permutation with $n$ elements for each $1\le n\le N$. This number is quite small for $N=50$ (it ends up being 1056) but grows quite rapidly. The number of permutations for a single LCM should be stored $\pmod{M-1}$ because $a^{M-1}\equiv 1\pmod{M}$ for all $0<a<M$ (Fermat's Little Theorem).

Subtask $N\le 500$:

In general, we can calculate the period of a permutation as follows:

So we can essentially solve the problem independently for each distinct prime power $D$. Doing this in $O(N^2)$ for a single prime power is sufficient for this subtask.

Subtask $N\le 3000$:

If we can count the number of ways to create a permutation of length $n$ for each $n\in [1,N]$ such that no cycle length is divisible by $D$ in $O\left(\frac{N^2}{D}\right)$ time, then the solution runs in $O\left(N^2\cdot \sum_{D=1}^N\frac{1}{D}\right)=O\left(N^2\log N\right)$ time.

Subtask $N\le 7500$:

The high bound on $N$ (hopefully) ensures that the above solution does not receive full credit. How can we do better?

Note that if $2D>N$ then we can compute the number of permutations containing a cycle with length divisible by $D$ in $O(1)$ time (assuming that we have precomputed some quantities in $O(N^2)$). This is true because if there is a cycle with length divisible by $p^k$, then there must be exactly one cycle with length equal to $p^k$ (and the rest can have arbitrary lengths).

Let's try to generalize. Define $D=p^k$. Any permutation has between $0$ and $\left\lfloor\frac{N}{D}\right\rfloor$ cycles with length divisible by $D$. So it suffices to count each of the following quantities for each $k\in \left[0,\left\lfloor\frac{N}{D}\right\rfloor\right]$.

If we can count both of these in $O\left(\frac{N^2}{D^2}\right)$ time, then this solution runs in $O\left(N^2\cdot \sum_{D=1}^{\infty}\frac{1}{D^2}\right)=O\left(N^2\cdot \frac{\pi^2}{6}\right)=O\left(N^2\right)$ time.

Mark Chen's code: 

#include <bits/stdc++.h>
using namespace std;
 
typedef long long LL;

int n; LL m;
 
typedef unsigned long long ull;
typedef __uint128_t L;
struct FastMod {
    ull b, m;
    FastMod(ull b) : b(b), m(ull((L(1) << 64) / b)) {}
    ull reduce(ull a) {
        ull q = (ull)((L(m) * a) >> 64);
        ull r = a - q * b; // can be proven that 0 <= r < 2*b
        return r >= b ? r - b : r;
    }
};
FastMod f(2);
 
LL mul(LL x, LL y) {
    return f.reduce(x * y);
}
 
LL add(LL x, LL y) {
    x += y;
    if (x >= m) x -= m;
    return x;
}
 
LL sub(LL x, LL y) {
    x -= y;
    if (x < 0) x += m;
    return x;
}
 
LL powmod(LL a, LL b) {LL res=1; a %= (m+1); for(;b;b>>=1) {if (b&1) res = res*a % (m+1); a = a*a % (m+1);} return res;}
 
const int MAXN = 7505;
LL factorial[MAXN], c[MAXN][MAXN];
 
int main() {
    freopen("exercise.in","r",stdin);
    freopen("exercise.out","w",stdout);
    cin >> n >> m;
 
    m--;
    f = FastMod(m);
 
    factorial[0] = 1;
    for (int i = 1; i < MAXN; ++i) factorial[i] = mul(factorial[i-1], i);
 
    c[1][0] = c[1][1] = 1;
    for (int i = 2; i < MAXN; i++) {
        c[i][0] = c[i][i] = 1;
        for (int j = 1; j < i; j++) c[i][j] = add(c[i-1][j-1], c[i-1][j]);
    }
 
    vector<int> composite(MAXN);
 
    LL ans = 1;
 
    for (int i = 2; i <= n; i++) {
        if (!composite[i]) {
            for (int j = i; j <= n; j *= i) {
                // count permutations of length j*k where ALL cycles are divisible by j
                vector<LL> aj(n/j+1);
                aj[0] = 1;
 
                for (int k = 1; k < n/j+1; k++) {
                    for (int l = 1; l <= k; l++) {
                        aj[k] = add(aj[k], mul(mul(c[j*k-1][j*l-1], factorial[j*l-1]), aj[k-l]));
                    }
                }
 
                // count permutations of length n-j*k where NO cycle is divisible by j
                vector<LL> nj(n/j+1);
 
                for (int k = n/j; k >= 0; k--) {
                    nj[k] = factorial[n-j*k];
                    for (int l = k+1; l <= n/j; l++) {
                        nj[k] = sub(nj[k], mul(c[n-j*k][n-l*j], mul(aj[l-k], nj[l])));
                    }
                }
 
                ans = (ans * powmod(i, sub(factorial[n], nj[0]))) % (m+1);
            }
 
            for (int j = 2*i; j <= n; j += i) {
                composite[j] = 1;
            }
        }
    }
 
    printf("%lld\n", ans);
}

My code (which uses the principle of inclusion and exclusion): 

#include <bits/stdc++.h>
using namespace std;
 
void setIO(string s) {
    ios_base::sync_with_stdio(0); cin.tie(0); 
    freopen((s+".in").c_str(),"r",stdin);
    freopen((s+".out").c_str(),"w",stdout);
}
 
typedef long long ll;
const int MX = 7501;
 
typedef unsigned long long ul;
typedef __uint128_t L;
struct ModFast {
    ul b, m; ModFast(ul b) : b(b), m(ul((L(1)<<64)/b)) {}
    ul reduce(ul a) {
        ul q = (ul)((L(m)*a)>>64), r = a-q*b;
        return r>=b?r-b:r; }
};
 
ModFast MF(1);
 
int M,MOD,n;
int ad(int a, int b) { 
    a += b; if (a >= M) a -= M;
    return a;
}
int sub(int a, int b) { 
    a -= b; if (a < 0) a += M;
    return a; 
}
int mul(int a, int b) { return MF.reduce((ul)a*b); }
 
int choose[MX][MX];
int with(int z) { // # of permutations with z dividing some cycle length
    int res = 0;
    vector<int> dp(n/z+1); dp[0] = sub(0,1);
    for (int i = 1; i <= n/z; ++i) for (int j = 1; j <= i; ++j)
        dp[i] = sub(dp[i],mul(choose[i*z-1][j*z-1],dp[i-j]));
    for (int i = 1; i <= n/z; ++i) 
        res = ad(res,mul(choose[n][n-i*z],dp[i]));
    return res;
}
 
ll mpow(ll a, ll b) { return !b?1:mpow(a*a%MOD,b/2)*(b&1?a:1)%MOD; }
 
int main() {
    setIO("exercise"); 
    cin >> n >> MOD; M = MOD-1; MF = ModFast(M);
    for (int i = 0; i <= n; ++i) {
        choose[i][0] = 1;
        for (int j = 0; j < i; ++j) choose[i][j+1] = mul(choose[i][j],i-j);
    }
    vector<bool> comp(n+1); ll ans = 1;
    for (int i = 2; i <= n; ++i) if (!comp[i]) {
        for (int j = 2*i; j <= n; j += i) comp[j] = 1;
        for (int j = i; j <= n; j *= i) ans = ans*mpow(i,with(j))%MOD;
    }
    cout << ans << "\n";
}

It is possible to solve this problem more quickly if you are familiar with exponential generating functions (EGF). This comment gives an explicit formula, which I'll try to explain here. From KACTL, we have the following fact. Let $g_S(n)$ be the number of $n$-permutations whose cycle lengths all belong to the set $S$. Letting $S(x)=\sum_{n\in S}\frac{x^n}{n}$, it follows that

$$\sum_{n=0}g_S(n)\frac{x^n}{n!}=\exp(S(x))=\sum_{k=0}^{\infty}\frac{S(x)^k}{k!}.$$
Essentially, the $\frac{S(x)^k}{k!}$ term corresponds to the number of ways to form permutations with exactly $k$ cycles. When all cycle lengths are valid,
$$S(x)=\sum_{n=1}^{\infty}\frac{x^n}{n}=-\ln(1-x)$$
and
$$\exp(S(x))=\frac{1}{1-x}=1+x+x^2+\cdots,$$
which is clearly correct. If we want to exclude cycles with lengths that are multiples of $z$, then
$$S(x)=-\ln(1-x)-\frac{1}{z}\cdot \sum_{k=1}^{\infty}\frac{x^{zk}}{k}=-\ln(1-x)+\frac{1}{z}\cdot \ln(1-x^z).$$
It follows that
$$\exp(S(x))=\frac{(1-x^z)^{1/z}}{1-x}.$$
By the binomial theorem, the numerator has only terms with degree divisible by $z$. Letting $d=\lfloor n/z\rfloor$ and $[x^n]P(x)$ denote the coefficient of $x^n$ in $P(x)$, it follows that
$$[x^n]\frac{(1-x^z)^{1/z}}{1-x}=[x^{zd}]\frac{(1-x^z)^{1/z}}{1-x^z}=[x^{zd}](1-x^z)^{1/z-1}.$$
By the binomial theorem,
$$g_S(n)=n!\cdot (-1)^d\binom{1/z-1}{d}=\frac{n!}{d!}\prod_{i=1}^d(i-1/z)=\frac{n!}{d!z^d}\cdot \prod_{i=1}^d(zi-1).$$
So it turns out that we can replace part of the above code with 

int without(int z) {
    int res = 1; 
    for (int i = 1; i <= n; ++i) {
        if (i%z != 0) res = mul(res,i);
        else res = mul(res,i-1);
    }
    return res;
}
int with(int z) { // # of permutations with z dividing some cycle length
    return sub(choose[n][n],without(z));
}

although this isn't actually faster since it still runs in $O(N \pi(N)) = O(N^2 / \log N)$ time (where $\pi(N)$ denotes the number of primes that are at most $N$).

Here is a way to derive this formula without generating functions (by Sanjeev):

(BEGIN)

We use $(n)_k$ to denote $n \cdot (n-1) \cdots (n-k+1)$, the falling factorial.

Let $a_n$ be the number of permutations that have no cycles with length dividing $z$. Then, if we imagine choosing the rest of the cycle that $1$ belongs to then recursing, we have

$$ a_n = \sum_{\substack{k = 1 \\ z \nmid k}}^n (n-1)_{k-1} a_{n-k}. $$
Expressing this in terms of $a_{n-z}$, we have
$$ a_n = (n-1)_z a_{n-z} + \sum_{k=1}^{z-1} (n-1)_{k-1} a_{n-k}. $$

You can think of the above as two cases: we choose a cycle of length greater than $z$ or less than $z$.

Now consider the corresponding expression for $a_{n-1}$:

$$ a_{n-1} = (n-2)_z a_{n-z-1} + \sum_{k=1}^{z-1} (n-2)_{k-1} a_{n-k-1}. $$

If we subtract $n-1$ times this expression from the expression for $a_n$, we get

$$\begin{align*} a_n &= (n-1)_z a_{n-z} + (n-1) \Big[a_{n-1} - (n-2)_z a_{n-z-1} - (n-2)_{z-2} a_{n-z}\Big] + (n-1)_0 a_{n-1} \\ &= n a_{n-1} + (n-1)_{z-1} (n-z-1) a_{n-z} - (n-1)_{z+1} a_{n-z-1}. \\ \end{align*}$$

This already implies an $O(n\pi(n))$ algorithm for the original problem after pre-computation, but we can do better. Manipulating the above, we see

$$\begin{align*} b_n \triangleq na_{n-1} - a_n &= (n-1)_{z-1} (n-z-1) \Big[ (n-z) a_{n-z-1} - a_{n-z}\Big] \\ &= (n-1)_{z-1} (n-z-1) b_{n-z}. \end{align*}$$

From the initial conditions, we see that $b_n$ is only nonzero when $z \mid n$. It is then straightforward by induction that $b_n = a_{n-1}$ when $z \mid n$, so we have

$$ a_n = \begin{cases} n a_{n-1} & \text{if } z \nmid n \\ (n-1) a_{n-1} & \text{else} \end{cases} = \frac{n!}{z^{\left\lfloor n/z\right\rfloor} \left \lfloor n/z\right\rfloor!} \prod_{i=1}^{\left\lfloor n/z\right\rfloor} (zi - 1) $$

If we precompute $\frac{n!}{z^{\left\lfloor n/z\right\rfloor} \left\lfloor n/z\right\rfloor!}$ for all prime powers $z \le n$ (noting that it is an integer), then after that we have an

$$ O\left(\sum_{\substack{1 \le z \le n, \\ z = p^k}} \frac{n}{z} \right) = O(n \log \log n) $$
algorithm for the problem. First factorize $M-1$ in $O(\sqrt{M})$ time (or $O(n)$, since prime factors of $M-1$ greater than $n$ do not affect our answer). Then we can do this precomputation in $O(n \log M)$ time by looping in increasing order of $z$ and keeping track of the powers of the various prime factors of $M-1$ in $\frac{n!}{\left\lfloor n/z\right\rfloor!}$, in addition to the part of it sharing no prime factors with $M-1$. For each prime power $z = p^k$, if it is coprime to $M-1$, we simply multiply $\frac{n!}{\left\lfloor n/z\right\rfloor!}$ by $z^{-\left\lfloor n/z\right\rfloor} \mod M-1$. Otherwise, we subtract $k \left\lfloor n/z\right\rfloor$ from the exponent of $p$ we have been keeping track of. Our final time complexity (assuming a word size of $\Omega(\max(\log n, \log M))$) is then

$$ O\left(n (\log \log n + \log M)\right), $$

and we require $O(n)$ space. The space can probably be improved to $O(n / \log n)$. Note that the constant factor for $\log M$ is favorable, since it comes from the maximum number of primes dividing $M-1$ (e.g. $9$ for $M\le10^9$).

(END)

Another solution that runs in $O(N\log N)$ time is to use divide and conquer to initialize a data structure that allows you to query any range product $l\cdot (l+1)\cdots (r-1)\cdot r$ modulo $M-1$ in constant time (where $1\le l\le r\le N$). This avoids the need to factorize $M-1$.