Containers for Video Reenactment as Inductive Bias for Content-Motion Disentanglement