feat: optional BM25+vector hybrid entity seeding with RRF fusion

[BigStartByXuyb]

Sorry, my English isn't very good. The following introduction was translated with the help of Claude. The commits were merged into one.

Summary

Adds an optional BM25 keyword index over entity names as a parallel seed path beside vector search in _get_node_data. Targets the exact failure mode in #3198: jargon/acronym queries where dense embeddings fail to seed the knowledge graph.

• Feature is off by default — zero behavior change unless enable_bm25_seeding=True
• Any BM25 path failure silently falls back to vector-only (a query never fails because of this)
• bm25s is an optional dependency (lazy import, graceful degradation)

What's Changed

Component	Description
BaseKeywordStorage	New pluggable storage interface (5th role)
Bm25KeywordStorage	Default impl: bm25s + TiktokenTokenizer (BPE, CJK-safe)
fuse_seed_rankings	RRF fusion of vector + BM25 ranked lists (k=60)
Query integration	QueryParam.enable_bm25_seeding flag, threaded through query chain
Lifecycle	Index rebuilt on insert + entity delete; workspace-isolated
Server config	LIGHTRAG_KEYWORD_STORAGE + ENABLE_BM25_SEEDING env vars

Design Decisions

• Entity names only (not chunks): smallest index that solves the seeding problem, avoids IDF drift
• Full rebuild (not incremental): milliseconds at entity-name scale, idempotent
• Fusion before graph expansion: one traversal on merged seeds, not two separate retrievals
• Local path only (MVP): global path is a follow-up

Test Plan

• 25 unit tests covering interface, BM25 build/search/persist/degrade, RRF fusion, flag-off regression, exception fallback, workspace isolation, CJK
• ruff check . clean
• QueryParam().enable_bm25_seeding == False confirmed

Closes #3198