Fix VLM models in refinement_loop.py: use actual vision models with correct providers

Add agentic refinement loop module