Teaching LLMs for Step-Level Automatic Math Correction via Reinforcement Learning