perf(proximity): broaden SIMD acceleration and remove reranking copies

[forhappy]

Summary

• Add runtime-dispatched AVX-512, AVX2, and NEON kernels for exact and scalar-quantized scoring.
• Preserve scalar-order f64 reductions and bit-identical scoring semantics.
• Score validated encoded vectors directly, removing intermediate reranking scratch copies across native, HNSW, PQ, composite, sync, and async paths.
• Reduce HNSW node cloning with Arc and use a lookup-only HashMap for native score caching.
• Add repeat, cache-reset, and SIMD-first benchmark controls.

Performance evidence

Preliminary local release measurements on an Apple M2 Max:

• Adaptive SQ8 at 1,536 dimensions improved from approximately 3.65–3.83 ms to 3.28–3.52 ms; logical work stayed at 1,005 quantized evaluations.
• HNSW at 128 dimensions improved from approximately 0.90–0.95 ms to 0.72–0.75 ms in the same harness; node reads and distance evaluations stayed unchanged at 194 and 212.
• Exact-search result counts and scalar/SIMD score bits remained unchanged.

These are preliminary one-shot measurements; the benchmark now supports repeated runs, cold/warm cache comparison, and randomized kernel order for more robust follow-up measurements.

Validation

• cargo fmt --all -- --check
• cargo check --target x86_64-apple-darwin --lib --tests
• cargo clippy --all-features --lib -- -D warnings
• Focused all-feature proximity tests: 23 passed
• cargo bench --all-features --no-run

The full all-feature test invocation remains blocked by pre-existing unresolved imports in tests/async_store.rs (catalog_map_id, control_record_key, control_root_name, ActiveIndexControl, and IndexControl). No persisted vector format changes are introduced.