Beyond Accuracy: Dissecting Mathematical Reasoning for LLMs Under Reinforcement Learning