Distributional Black-Box Model Inversion Attack with Multi-Agent Reinforcement Learning