High Performance Simulation for Scalable Multi-Agent Reinforcement Learning