feat: add first-class local Ollama review

[EmersonBraun]

What changed

• adds native Ollama chat and tool-call handling
• wires the local provider into the CLI
• documents a bounded local-review recipe and operational safeguards
• refreshes machine-readable documentation
• adds adapter and end-to-end CLI tests using a synthetic local endpoint

Why

Teams can run code review with local inference instead of sending repository content to a hosted model provider. The recipe keeps model choice, file limits, concurrency, failure behavior, and network boundaries explicit.

Validation

• npm run check — 21 tests passed
• npm pack --dry-run — 43 files
• git diff --check
• real local smoke test with Ollama 0.21.0 and qwen3:4b confirmed model loading, stdin ingestion, and successful /api/chat calls without credentials or external source code

The real smoke test also showed that qwen3:4b is slow for the complete seven-lens sequential review. The documented qwen2.5-coder:7b recipe remains the recommended practical baseline; this is a model-performance constraint, not an adapter failure.

This PR does not include an upstream Ollama contribution, release, tag, or package publication.