DoughNet: A Visual Predictive Model for Topological Manipulation of Deformable Objects