Improve BLS performance: cache parsed keys, add batch verification, add benchmarks

Summary

All BLS operations take raw byte arrays (types.rs#L17-L21), so every call re-deserializes (verify does a G1 decompression per call) and re-runs subgroup checks — verify passes pk_validate = true and sig_groupcheck = true on every invocation. For a DV node the public shares are cluster-constant; nothing is memoized.

There is also no batch verification: blst's verify_multiple_aggregate_signatures is unused, and the hot paths verify one signature at a time — every incoming peer partial signature (parsigex → verify_eth2_signed_data → signing.rs#L148), a nested validators×partials loop in dkg/src/aggregate.rs#L146, and one builder registration per validator in cluster/src/lock.rs#L427.

Additionally, aggregate_public_keys hand-rolls point addition through unsafe FFI and skips per-key subgroup validation — blst's safe AggregatePublicKey::aggregate(..., validate) covers it. A second independent blst wrapper with safe constructors already exists in pluto-frost (crates/frost/src/curve.rs); the ad-hoc scalar helpers in blst_impl.rs duplicate it.

There are no benchmarks: the only criterion setup in the workspace is crates/k1util/benches, while Charon benchmarks its BLS backend (tbls_test.go BenchmarkHerumiImplementation).

Proposed change

• Introduce parsed key/signature types (wrapping blst types) validated once at construction; keep raw bytes only at serialization boundaries.
• Add a batch-verification entry point and use it on the partial-signature paths.
• Replace the unsafe aggregation with the blst safe API; consolidate scalar helpers with pluto-frost's wrappers.
• Hoist the O(n²) scalar_from_u64 recomputation out of the interpolation loops — evaluate_polynomial rebuilds the same scalar every iteration and compute_lagrange_coefficients reconstructs per-index scalars inside the inner loop; build them once per index.
• Add criterion benches (sign/verify/aggregate/threshold_aggregate) using the k1util bench as a template, so wins are measurable.

Start with crates/crypto/src/blst_impl.rs and trace the verification paths through crates/eth2util/src/signing.rs, dkg/src/aggregate.rs, and cluster/src/lock.rs. Use crates/k1util/benches as the benchmark template and review crates/frost/src/curve.rs for existing wrappers. Done means the proposed caching, batch verification, safe aggregation, scalar-hoisting, and criterion benchmarks are implemented and measurable.