Model-Based Reinforcement Learning via Latent-Space Collocation