feat(rag): implement Multi-Query Expansion for BM25 search

[Xenon010101]

Summary

Adds Multi-Query Expansion to improve BM25 retrieval in the RAG pipeline. The module generates paraphrased query variants via the LLM, runs BM25 search for each, and merges results using Reciprocal Rank Fusion (RRF).

Changes

backend/app/rag/multi_query.py (new)

• BM25 class — Pure-Python BM25Okapi implementation (no new dependencies)
• generate_query_variations() — Uses InferenceClient to generate 4 paraphrased query variants
• reciprocal_rank_fusion() — Merges multiple ranked result lists with RRF (k=60)
• multi_query_retrieve() — End-to-end pipeline: fetch chunks from ChromaDB, build BM25 index, generate variants, search each, RRF merge, top-K

backend/app/rag/retriever.py

• When MULTI_QUERY_ENABLED=True (default), Stage 1 uses multi_query_retrieve instead of the ChromaDB embedding search
• Falls back to the existing embedding search when multi-query is disabled
• Cross-encoder reranking (Stage 2) runs on the BM25 results as before

backend/app/config.py

• Added MULTI_QUERY_ENABLED: bool = True setting

How it works

1. User sends a query
2. LLM generates 4 paraphrased variants
3. Each variant searches the BM25 index built from the user's document chunks
4. Results are merged via RRF (deduplicated, reranked)
5. Top-10 results proceed to the cross-encoder reranker
6. Final top-5 are returned to the LLM for answer generation

Design decisions

• No new dependencies — BM25 implemented inline (~30 lines); uses existing huggingface_hub and chromadb packages
• Graceful degradation — If LLM call fails, only the original query is used; if no chunks exist, returns empty
• Configurable — Set MULTI_QUERY_ENABLED=false to restore original ChromaDB-only retrieval

Closes #283