Speed up polysemous training with AVX-512.

[mulugetam]

This change adds AVX-512 implementations of the compute_cost and cost_update functions used in polysemous training. Benchmarks of the training phase with benchs/bench_polysemous_sift1m.py show ~1.5x speedup over the existing scalar implementation, which relied only on partial auto-vectorization.

Once the SIMD support overhaul is complete, I hope this can be considered for integration.
cc: @mdouze @subhadeepkaran

[bshethmeta]

@mnorris11 @subhadeepkaran Do you have enough context to review this?

[subhadeepkaran]

@mnorris11 @subhadeepkaran Do you have enough context to review this?

Yep, you can assign it to me. the change can be reviewed and merged post dynamic dispatch landing