perf: Batch embeddings in RAGIndexer instead of per-file encode

[omsherikar]

Scope

refactron/rag/indexer.py — add_chunks / index_repository

Problem

Each file path through _index_file → add_chunks calls self.embedding_model.encode(documents, ...) for only that file's chunks. SentenceTransformer/GPU throughput is much better with larger batches (e.g. hundreds of texts per call) than many small calls.

Suggested direction

• Accumulate (chunks, documents) across files (with a max batch size / memory cap), then encode in batches and collection.add in corresponding batches.
• Optional: configurable --batch-size for index CLI.

Acceptance

• Benchmark: indexing N files shows fewer encode invocations and lower wall-clock time on representative repos (document rough numbers in PR).