ReaPER: Improving Sample Efficiency in Model-Based Latent Imagination