- When model starts with 'ollama:', construct api_base from request host and port
- api_base is now passed to LiteLLMBackend for local connections