Text this: Data-efficient reinforcement learning with self-predictive representations