p256 acceleration

[austinabell]

Currently accelerates the muls and is ~8-9x speedup, but there is a lot of slop due to the unaccelerated conversion from projective to affine. Also unfortunate that this implementation has BE representation internally, so some unfortunate copies again as with k256.

TODO:

• Accelerate proj -> affine -> proj
• Reduce conversion logic
  • Unfortunate this is so generic, and there might be some extra copies that can't be guaranteed to be optimized away
• Explore different bit widths acceleration (out of scope for this)
  • Currently just errors in compile time with this (which I think is correct, shouldn't accidentally be using this for other crates in this repo
• Add acceleration for other code paths other than just mul (not sure where/how it might be used)
• Check const time signature gen (or panic on usage in zkvm) documented non const time
• Update version to 1.2.1 when released (requires Add modular arithmetic precompiles risc0#2653 to be included)

[tzerrell]

I'm not convinced we're tracking the identity correctly, especially in various places we make affine/projective conversions and conversions in and out of risc zero types. Can you take a look and either make fixes or help me understand how the identity is being handled?

[tzerrell]

Thanks for the updates, they look good!