Learning to solve hard problems in RL for LLMs by never giving up