Compute- and Memory-Efficient Reinforcement Learning with Latent Experience Replay