Learning to Plan Optimistically: Uncertainty-Guided Deep Exploration via Latent Model Ensembles