EUCLID: Towards Efficient Unsupervised Reinforcement Learning with Multi-choice Dynamics Model