Bootstrapped Representations in Reinforcement Learning