On Uncertainty in Deep State Space Models for Model-Based Reinforcement Learning