Vision-Language Models Provide Promptable Representations for Reinforcement Learning