Learning Invariant Representations for Reinforcement Learning without Reconstruction