Fix: lazy model load on first request, add error handling

Fix: replace llama_cpp.server with custom FastAPI server (no mlock, no health check loop)