M1.5.3: Implement Ollama adapter (local LLM)

[LyeZinho]

Use Ollama HTTP API client, implement model loading and caching, implement streaming response handling, add fallback to Gemini if Ollama unavailable, write tests with local Ollama instance