Rich-Observation Reinforcement Learning with Continuous Latent Dynamics