Fault-Tolerant Federated Reinforcement Learning with Theoretical Guarantee