Distributed vector search

[jasonstack]

• query all replicas selected by consistency level at once with full request range from ReadCommand

• filter top-K results at coordinator in QueryPlan#postProcessor using PriorityQueue and discard rows with the lowest scores when exceeding limit

• fail vector search if limit exceeds MAX_TOP_K (default 1000) to avoid overflow HNSW data structure or OOM in QueryPlan#postProcessor

• skip short-read-protection, read repair and replica filtering protection because replica response will be top-k. Data mismatch may be caused by top-k filtering instead of actual data inconsistency.

• paging will result in lower recall, because coordinator resumes from last returned row's partition

TODO: replica should return top-k result to coordinator