perf(map): Optimize LLM pipeline with merged prompts, model tiering, and async

[JoshDoesIT]

Summary

Reduce total LLM calls from ~573 to ~100-150 per run through three phases of optimization, targeting both the map enrichment pipeline and the parse pipeline.

Changes

Phase 1 — Concurrency Foundation

• Add async LLM client methods (call_llm_async, generate_async, etc.) using ollama.AsyncClient
• Add SQLite-backed LLM response cache (LLMCache in cache.py)
• Add query_by_embedding() for batch vector queries
• Batch embedding in mapper via embed_batch()
• Add --concurrency and --cache CLI options to ctrlmap map

Phase 2 — Call Reduction

• Merged relevance+rationale prompt — combines two separate LLM calls into one via merged_relevance_rationale.txt and evaluate_chunk_async(), eliminating ~200 calls
• Streaming per-control pipeline — each control flows through all steps independently instead of waiting for all controls per step
• Raise min_score from 0.35 → 0.50 to skip weak matches before they reach the LLM

Phase 3 — Model Tiering & Remaining Optimizations

• Model tiering — qwen2.5:7b for simple tasks (meta-classify, gap rationale, parse), qwen2.5:14b for accuracy-critical compliance evaluation (~2x faster for 60% of calls)
• Async parse pipeline — llm_chunker.py now uses asyncio.gather() for concurrent page extraction (3-5x parse speedup)
• Cache wired into call_llm_async() — transparent get/put makes --cache flag actually work for near-instant re-runs
• Async gap rationale — true async instead of sync-in-async wrapper
• Top-K 10 → 5 — fewer chunks sent to LLM evaluation
• Embedder singleton — @functools.cache shares the SentenceTransformer model across pipeline stages
• Skip meta-classify for mapped controls — ~20 fewer LLM calls

Test Plan

• All existing tests pass (uv run pytest) — 229/229
• New tests written following TDD (Red → Green → Refactor)
• Linting passes (uv run ruff check .)
• Type checking passes (uv run mypy src/)
• Eval suite: relevance/compliance/meta at 100%, 2 pre-existing faithfulness flakes unchanged

Checklist

• Tests written before implementation (TDD)
• Documentation updated (ARCHITECTURE.md)
• No unused code, console logs, or dead comments
• Follows existing project coding style