HNSW vector search

Vector fields are stored per segment and now get an HNSW graph at flush time by default. VectorQuery uses the graph when it is present, then reranks the shortlist with exact cosine similarity.

The important knobs are:

Setting	Use
BuildHnswOnFlush	Disable only for tiny indices or controlled comparisons
NormaliseVectors	Keep on for cosine search unless vectors are already prepared
HnswBuildConfig.M	Higher values improve recall and increase index size
VectorQuery.EfSearch	Higher values improve recall and increase latency
OversamplingFactor	Rerank a larger shortlist for better final ordering

Filters are handled by selectivity. Very tight filters scan the matched docs. Medium filters use an allow-list during graph traversal. Loose filters use post-filtering with retries.

See Vector search.