feat: add LLM enrichment for model name extraction

[cassiocouto]

Summary

Closes #36 — adds optional LLM-based enrichment to extract specific model names (e.g., gpt-4o, claude-3-opus-20240229) from code snippets around detected AI components.

• New --llm-enrich CLI flag with --llm-model, --llm-api-key, and --llm-base-url options
• Uses litellm as a unified LLM client — supports OpenAI, Anthropic, Ollama (local), and 100+ providers
• Only enriches llm_provider and model component types with empty model_name (filters out containers, tools, MCP servers, etc.)
• Reads ~20 lines of source context around detection sites for accurate extraction
• Cross-references extracted names against the built-in model registry for provider/deprecation metadata
• Batched LLM calls (default 5 per request) with graceful fallback to individual calls on error
• Dependency-missing guard: clear error with install hint when litellm is not installed
• Privacy warning for non-local models; recommends ollama/* for sensitive codebases
• 39 new tests (unit + CLI integration), all mocked — no real LLM calls in tests
• New docs/enrichment.md with usage guide, privacy section, and cost guidance
• Also fixes two pre-existing test failures (test_demo_command, test_sarif_relative_path_calculation on Windows)

Test plan

• pytest tests/test_enrichment/ — 39 tests pass (prompt templates, JSON parsing, component filtering, batch/single enrichment, error handling, CLI flags, privacy warnings)
• pytest tests/ — full suite: 786 passed, 0 failed
• ruff check — no lint errors on new/modified files
• Manual test with --llm-enrich --llm-model ollama/llama3 against a real project (requires Ollama running locally)