Learning from Failures in Multi-Attempt Reinforcement Learning