A Max-Min Entropy Framework for Reinforcement Learning