Data-Efficient Reinforcement Learning with Self-Predictive Representations