Learning Temporally-Consistent Representations for Data-Efficient Reinforcement Learning