[feat] shared base MSM many

[jonathanpwang]

An untested hypothesis is that when the number of columns is high, the current version of doing best_multiexp one column at a time is not optimal because it is not memory local and doesn't use the fact that the base of the MSM is always the same (namely the KZG params).

This is a quick edit so that for a batch of columns, when we commit_lagrange we do all MSM for a chunk of the bases in a single thread.

Haven't tested it on any wide circuits to compare.

[jonathanpwang]

edit: tried keccak_bits and on my laptop (Arm, better memory) it is worse than main which just uses rayon