Selective Perception: Optimizing State Descriptions with Reinforcement Learning for Language Model Actors