SSR: Socratic Self-Refine for Large Language Model Reasoning