MIR2: Towards Provably Robust Multi-Agent Reinforcement Learning by Mutual Information Regularization