Visual Imitation Learning of Non-Prehensile Manipulation Tasks with Dynamics-Supervised Models