PerSim: Data-efficient Offline Reinforcement Learning with Heterogeneous Agents via Personalized Simulators