Learning a Multi-Agent Simulator from Offline Demonstrations