Training Large Language Models for Reasoning through Reverse Curriculum Reinforcement Learning