perf(sumcheck): replace par_bridge with par_chunks/par_iter (-7.1% big)

[tamirhemo]

Summary

• Replace par_bridge() with par_chunks_mut() / into_par_iter() in three hot-path files
• par_bridge uses a mutex internally to pull from sequential iterators, causing lock contention
• The replacements use rayon's native work-stealing without a mutex

Changed files

• slop/crates/multilinear/src/restrict.rs — mle_fix_last_variable
• slop/crates/jagged/src/poly.rs — eval and partial_jagged_little_polynomial_evaluation
• crates/hypercube/src/prover/zerocheck/sum_as_poly.rs — zerocheck sum accumulation

Benchmark results (on top of E1 + E2)

Workload	Before	After	Delta
fib	26,504ms	24,694ms	-6.8%
keccak	34,845ms	32,207ms	-7.6%
big	37,899ms	35,215ms	-7.1%

Cumulative improvement (E1+E2+E8 vs original baseline)

Workload	Original	Current	Total Delta
fib	35,380ms	24,694ms	-30.2%
keccak	42,847ms	32,207ms	-24.8%
big	47,697ms	35,215ms	-26.2%

Stack

1. feat(bench): add CPU prover benchmark harness and baseline profiling #2726 — bench infra
2. perf(prover): default rayon to available_parallelism #2727 — E1: physical cores (-19.8%)
3. perf(bench): add optional mimalloc allocator for sp1-perf (-7.2% fib) #2728 — E2: mimalloc (-7.2% fib)
4. This PR — E8: remove par_bridge (-7.1%)

Test plan

• cargo test --release -p slop-multilinear — 0 tests (no test coverage for restrict.rs)
• cargo test --release -p slop-jagged — 13 passed
• cargo test --release -p sp1-hypercube — 14 passed
• cargo test --release -p sp1-prover test_e2e_node — full e2e (long)

🤖 Generated with Claude Code