perf(vector): parallelize large IVF scans

[ppodolsky]

Summary

• parallelize large IVF-TQ and binary-IVF scans with bounded worker-local tasks
• cover both distinct top-k retrieval and multi-value combined document scoring
• preserve serial execution for small scans, single-thread pools, and WASM
• add forced-parallel correctness tests plus ignored reproducible benchmarks

Benchmarks

Release mode on 8 Rayon threads:

• IVF-TQ distinct, 131072 postings, dim 128: 1.042 ms -> 0.460 ms (2.26x)
• IVF-TQ combined, 131072 postings, dim 128: 4.397 ms -> 1.902 ms (2.31x)
• Binary IVF distinct, 131072 postings, 2048 bits: 0.800 ms -> 0.465 ms (1.72x)
• Binary IVF combined, 131072 postings, 2048 bits: 1.463 ms -> 1.304 ms (1.12x)

At the 65536-posting threshold, all four paths remain faster than forced serial execution. Binary combined gains are smaller because document aggregation and sorting dominate its runtime.

Validation

• cargo fmt --all -- --check
• cargo clippy -p hermes-core --features native --all-targets -- -D warnings
• cargo clippy -p hermes-core --no-default-features --features native --all-targets -- -D warnings
• RUSTFLAGS=-Dwarnings cargo check -p hermes-core --no-default-features --features wasm
• cargo test -p hermes-core --lib (1051 passed, 9 ignored)