The Benefits of Power Regularization in Cooperative Reinforcement Learning