feat(rag): add LLM-based pointwise reranker to RAG pipeline

[pane2004]

Adds an LLM-based pointwise reranker to the RAG pipeline, significantly improving retrieval quality.

Key Features

• Pointwise scoring: Each passage is scored 1-10 based on relevance to the query using a detailed grading rubric
• Parallel batch processing: Passages are distributed round-robin across N batches and scored concurrently using asyncio.gather(): https://fin.ai/research/using-llms-as-a-reranker-for-rag-a-practical-guide
• Score thresholding: Passages scoring below 5 are automatically filtered out, reducing noise in results
• Graceful fallback: On timeout or LLM failure, falls back to original embedding scores

Why Reranking Matters

The demo shows a dramatic improvement - vector search alone ranked the two most relevant passages at positions 13 and 14, while irrelevant passages with keyword overlap ranked 1-3.
After LLM reranking:

• High-relevance passages → 1 and 2 with 10/10 scores
• All irrelevant passages → filtered out completely

Files Changed

New files:

• src/flare_ai_kit/rag/vector/reranker/ - New reranker module
  • base.py - Abstract BaseReranker class
  • gemini_reranker.py - GeminiPointwiseReranker implementation
  • prompts.py - Grading rubric and prompt templates
• examples/08_rag_reranker_demo.py - Full RAG + reranker demo
• tests/unit/rag/vector/reranker/ - 19 unit tests

Modified files:

• src/flare_ai_kit/rag/vector/settings.py - Added RerankerSettings
• src/flare_ai_kit/rag/vector/factory.py - Integrated reranker into VectorRAGPipeline
• src/flare_ai_kit/rag/vector/api.py - Exported new classes
• src/flare_ai_kit/common/exceptions.py - Added RerankerError, RerankerParseError

Test Plan

• All 19 unit tests pass
• Ruff formatting and linting pass
• Pyright type checking passes
• Manual testing with demo script shows correct reranking behavior

Closes #10