Leveraging Language Models for Out-of-Distribution Recovery in Reinforcement Learning