Bounded MSM

[mmagician]

Description

For our use case of polynomial with small coefficients, we are interested in doing MSMs over small scalars.

For MAX_BITS = 30 we have a massive 7x performance improvement:

MSM small for Bls12_381::G1    time:   [186.06 ms 188.58 ms 191.62 ms]

versus the naive MSM:

MSM for Bls12_381::G1          time:   [1.3238 s 1.3446 s 1.3680 s]

For MAX_BITS = 60, we have:

time:   [303.07 ms 304.08 ms 305.53 ms]

This is actually "only" 5x better. Surprisingly, or perhaps not*, this is MUCH better than claimed in Lasso: https://github.com/a16z/Lasso/blob/823da601422d17f7c150631947b33a9db1ad5b98/src/msm/README.md?plain=1#L2

*I think that the claim in Lasso actually refers to the costs of the entire scheme, instead of just MSM.

Also, the interface change is a little ugly VERY BAD, as the default is almost never respected.
Probably we'd need to also add this const to sw Projective and all its implementations... it's quickly turning messy, but I don't know if there's a better way.