Plan To Predict: Learning an Uncertainty-Foreseeing Model for Model-Based Reinforcement Learning