Saving the Limping: Fault-tolerant Quadruped Locomotion via Reinforcement Learning