nmod_mat_nmod_vec_mul, nmod_mat_nmod_vec_mul_ptr

The rows of $B$ are accumulated with delayed reductions where a vector code path exists: with AVX512-IFMA for $n \le 2^{52}$ (as in nmod_mat_mul_u52), with AVX2 or AVX-512 without IFMA for $n \le 2^{32}$ (splitting each product at bit 56 as in Splitting at bit 56; the next path is used instead for short vectors: $k \le 6$, or $k \le 32$ and $B$ has at least $\max(k, 16)$ columns), with a vector FMA for $n < 2^{50}$ (as in nmod_mat_mul_fp50), and with AVX2 or AVX-512 for the other moduli, by accumulating the products of limbs of at most 32 bits of the entries as in Split limbs (AVX2, AVX-512), with one reduction of each entry of $c$ at the end (Thin matrices). These code paths are not used when $B$ is too narrow for them, with thresholds on the length $k$ of $a$ and the number of columns that depend on the path (for instance, with AVX512-IFMA, from 16 columns for $k = 2$ down to 3 or 4 columns from $k = 6$ on; 5 columns for the moduli of the last path). Otherwise, with one column, or up to 8 columns and $4 \le k \le 32$ ($8 \le k \le 32$ for $n > 2^{62}$), each entry of $c$ is a dot product; else each row is added with _nmod_vec_scalar_addmul_nmod when $k$ is less than 6 (for $n \le 2^{32}$), 8 (up to $2^{62}$) or 10, or when $B$ has at least 2048 columns (4096 for $n \ge 2^{63}$), and the product is done with nmod_mat_mul_classical on blocks of rows of $B$ for longer vectors.