Compressed Conditional Mean Embeddings for Model-Based Reinforcement Learning