Adding embedding-based retrieval

[Shyamsaibethina]

• Adds a Retriever component that uses sentence-transformer embeddings to select the top-k most relevant playbook bullets per test sample, instead of passing the full playbook to the generator
• Retrieval is opt-in via --retriever_top_k. When omitted, testing uses the full playbook as before
• Tracks and reports average sub-playbook token count alongside accuracy