feat: add LlamaIndex integration with hybrid retriever support (#217)

[ArshadJafri]

Description

This PR implements LlamaIndex integration for LEANN, requested in #217 and addressing integration blockers discovered previously.

What was built

We introduced two BaseRetriever subclasses:

1. LeannRetriever: A seamless integration for pure semantic/vector search.
2. LeannHybridRetriever: A native hybrid retriever leveraging LEANN's built-in BM25 capabilities natively.

Key Fixes Included

• Parameter Safety: Accurately mapped the user-facing keyword bm25_weight (e.g., 0.3) to LEANN's underlying vector-weight parameter gemma.
• Tuple Normalization Fix (Support LEANN in llamaindex #217): Properly intercepts SearchResult objects and normalizes their metadata into strict dictionaries before casting them to LlamaIndex NodeWithScore objects. This natively prevents crashes when users wrap this inside LlamaIndex's QueryFusionRetriever.

Testing

• ✅ Tested bm25_weight -> gemma translation mathematics.
• ✅ Added visually verifiable local tests under examples/llamaindex_hybrid_example.py.

[ASuresh0524]

LGTM @ArshadJafri

Thanks for this integration — the implementation looks good.

Verified:

• bm25_weight → gemma mapping is correct (bm25_weight=0.3 → 70% vector, 30% keyword)
• Metadata normalization in _results_to_nodes fixes the QueryFusionRetriever crash from Support LEANN in llamaindex #217
• API usage: LeannSearcher.search() and gemma are used correctly
• Dependencies are already covered via llama-index-core in leann-core

Minor suggestions (before merging):

• Example: consider using a temp directory in examples/llamaindex_hybrid_example.py and cleaning up the index file afterward.
• Docs: A short note in the docs about this integration would help discoverability.