Optimize for sub-block transform lengths in fft_small transformed rings

[fredrik-johansson]

Transformed representations were previously limited to the fft_small block size granularity 256, but can now go down to length 32, 64 or 128.

As a result the input length crossover where FFT wins in nmod_poly_mat_mul roughly halves to ~32 and the cutoff is adjusted accordingly.

For fmpz_mat_mul_fft_small, the crossovers drop ~1000 bits, but I left the existing algorithm cutoffs for now since the effect isn't huge.

Done using Claude Fable 5.

I also tried to smoothen out the jumps at small sizes by adding transforms of length $3 \cdot 2^k$. This does speed up lengths 33-48 for nmod_poly_mat_mul by 5-20% but gives inconsistent speedups at lengths 65-96 or 129-192 with a ~10% slowdown for large matrix dimensions. Omitted from this PR to keep things simple, but may be worth resurrecting with better optimized transform kernels; see the attached patch

0003-composed-lengths.patch