bench: microbenches for the hot mixed base x extension field kernels (chunk 1 of #260)

[koko1123]

Chunk 1 of the delayed modular reduction plan tracked in #260.

What

Ignored-test microbenches (same pattern as crates/backend/koala-bear/src/benchmark_poseidons.rs) for the four hot mixed base x extension kernels that later chunks will retile:

• crates/backend/sumcheck/src/benchmark_product_sumcheck.rs -- product-sumcheck quadratic round (compute_product_sumcheck_polynomial over base-packed evals x extension-packed weights, i.e. the sumcheck_quadratic hot loop).
• crates/whir/src/benchmark_first_round.rs -- WHIR round-0 combine (combine_and_compute_first_round; made pub(crate) so the bench module can reach it, no other production change).
• crates/backend/poly/src/benchmark_kernels.rs -- eval_base_packed and finger_print_packed (memory-style 2-element and bytecode-style 13-element logup tuples).

Each bench first asserts the kernel output against an independent scalar reference computation (recursive evaluate_sequential for eval_base_packed, materialized eq-tables for the WHIR combine, plain scalar loops for the rest), so the harness doubles as a regression test when the kernels are rewritten in later chunks.

Run with e.g.:

cargo test --release --package poly --lib -- benchmark_kernels --nocapture --ignored
cargo test --release --package sumcheck --lib -- benchmark_product_sumcheck --nocapture --ignored
cargo test --release --package whir --lib -- benchmark_first_round --nocapture --ignored

Baseline numbers (Apple Silicon, NEON, current eager kernels)

product sumcheck quadratic round (20 vars): 0.416 ms/call, 2519 Melems/s
WHIR round-0 combine (20 vars, 4 statements): 3.884 ms/call, 270 Melems/s
eval_base_packed (22 vars): 1.105 ms/call, 3796 Melems/s
finger_print_packed (memory-style, 2 data, 16 alphas): 15.8 ns/call, 253M scalar rows/s
finger_print_packed (bytecode-style, 13 data, 16 alphas): 75.0 ns/call, 53M scalar rows/s

Checks

• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets -- -Dwarnings
• cargo test --release --all

Chunks 2+ (Algebra::mixed_dot_product, delayed packed-quintic Mul, call-site retiling, NEON dot_product_8) will follow as separate PRs, each gated on these benches plus the E2E XMSS baseline, per the plan in #260.

🤖 Generated with Claude Code