Probabilistic View of Multi-agent Reinforcement Learning: A Unified Approach