A Mixture of Surprises for Unsupervised Reinforcement Learning