[sync] upstream inverse acceleration

[Velaciela]

Description:

From snarkify merkle circuit example (comparison),
we find that the performance of the poseidon hash in the Scroll version has a gap compared to the upstream.

• MerkleTreeUpdateChip::new(proof.clone())
  • assert!(proof.verify());
    • hash(old, *right), hash(new, *right)
      • Spec::new()

And it turns out that the inverse() operation is the main bottleneck.

related PR

This PR cherry-picked several performance-optimizing PRs from upstream:

• Fast modular inverse - 9.4x acceleration privacy-scaling-explorations/halo2curves#83
• Fast isSquare / Legendre symbol / Jacobi symbol - 16.8x acceleration privacy-scaling-explorations/halo2curves#95

performance

before

Start: kzg k_table_size=18 repeat_count=10
··Start: keygen
··End: keygen............................................................175.00s
··Start: prove
··End: prove.............................................................106.00s
··Start: verify
··End: verify..............................................................0.01s
End: kzg k_table_size=18 repeat_count=10.................................290.00s

after

Start: kzg k_table_size=18 repeat_count=10
··Start: keygen
··End: keygen.............................................................40.40s
··Start: prove
··End: prove..............................................................38.80s
··Start: verify
··End: verify..............................................................0.01s
End: kzg k_table_size=18 repeat_count=10..................................89.10s