fft_small_plan_init_nmod

xtrunc_max bounds the operand lengths, rounded up to a multiple of 256. For a product of lengths $a_n$ and $b_n$, take len_bound $= \min(a_n, b_n)$ and prod_bits $= 2 \lceil \log_2 n \rceil$; for a sum of $K$ products, len_bound $= K \min(a_n, b_n)$. Unreduced inputs, such as sums of operands formed in transform space, are accounted for by increasing prod_bits. If $z_n \le z_l + 2^d$ for a power of two $2^d$ that is at least $z_h$ and the operand lengths, the transform has length $2^d$ and the excess coefficients wrap around below $z_l$. If mod.n is a context prime, one prime is used without Chinese remaindering. If direct_len is at least 100, $n$ has at least 20 bits, $2^d \mid n - 1$ and $n$ is prime, the transform is done modulo $n$ itself; pass 0 to disable this.