PMIC: Improving Multi-Agent Reinforcement Learning with Progressive Mutual Information Collaboration