batched bn254 51 bits RNE multiplier

[xrvdg]

• Add 51-bit floating-point Montgomery multiplication for BN254 scalar field optimised for WASM
• Rename block-multiplier to bn254-multiplier with reorganized module structure
• Optimize carries and squaring operations in the b51 multiplier
• Add Kani verification for integer-to-float conversion
• Enable relaxed SIMD in WASI runners
• Make benchmarks WASI-compatible

Reviewing: Not all file movements and renames are tracked properly by Github's diff view. The important folder is rne which contain the new Montgomery multiplication implementation. I've tried to stay true to the structure of the one in rtz.

[ocdbytes]

LGTM !
some nit picks and some personal doubts I had while reviewing.

Suggestion (Not necessary) : Also we can add edge case tests for 0 values, identity values etc. which may not come in proptest for simd_mul and simd_sqr.

[ocdbytes]

Doubt : why are we not checking the limits of the inputs here ?

[xrvdg]

The main reason is that these are internal functions to a kernel that runs a hot loop. The kernel needs to be designed in a way that the precondition holds because checking in a hot loop is costly. This note is there as a reminder to let the kernel have this invariant.

I did add a debug assert to u256_to_u255 to check if the invariant is actually being uphold when running in debug mode.

[ocdbytes]

cool works. Thanks for clarifying

[xrvdg]

Suggestion (Not necessary) : Also we can add edge case tests for 0 values, identity values etc. which may not come in proptest for simd_mul and simd_sqr.

I try to not have overlap in tests. They are mathematically special but they are not edge cases. I would have used those if we didn't have a reference to test against.

[ocdbytes]

LGTM !!

[Bisht13]

Nit: test_simd_sqr validates against simd_mul(a, a, b, b) but not independently against ark_ff_reference. Since test_simd_mul already validates mul against ark_ff, the chain sqr → mul → ark_ff provides transitive coverage. But the sqr-specific code path (cross-terms doubled + diagonal terms added separately) differs from mul's schoolbook, so a direct ark_ff_reference(a, a) check would add confidence for free:

let a2_ref = ark_ff_reference(a, a);
prop_assert_eq!(Fr::new(BigInt(a2s)), a2_ref);

[xrvdg]

simd_sqr can be bitwise compared against simd_mul without a transformation (happens within ark_ff_reference and Fr::new). This makes sqr-specific code path easier to debug. Property-based tests with transitivity gives us adequate coverage.