Adds scalar quantization support for HNSW vector indexes behind the quantization option.

This PR supports:

• quantization := 'sq8'
• quantization := 'sq16'
• quantized in-memory distance computation for ANN traversal
• optional full-precision reranking via use_full_precision_rerank
• SIMD-friendly SQ16 dot-product kernel with scalar fallback
• quantized query-vector/cache alignment fixes
• visibility checks for quantized on-disk embeddings
• quantized cache invalidation on index mutation

Dot product with quantization is rejected for now because the quantized metric path currently supports
L2/L2SQ/Cosine-style scoring.

Behavior

Without quantization, HNSW continues to use the existing full-precision float embedding path.

With quantization enabled:

• index traversal uses quantized embeddings
• query vectors are quantized once per query
• cached quantized payloads store compact integer vectors plus scale/norm metadata
• if use_full_precision_rerank := true, final candidates are rescored using full-precision embeddings
• if use_full_precision_rerank := false, results are returned directly from the quantized metric path

Benchmark

Ran it on OpenAI 50k small

Results:

Speedup vs baseline:

Storage overhead vs baseline:

Correctness Notes

This PR also fixes several issues found during testing:

• quantized query buffer sizing/alignment
• SQ8/SQ16 cached metadata layout
• traversal using the quantized metric function instead of the full-precision metric
• stale quantized cache after index mutation
• visibility checks for quantized on-disk embeddings

Quantized cache invalidation is currently coarse-grained: mutations invalidate the full quantized cache.
A future improvement should make this transaction-aware and row/version based.