MoCoGAN: Decomposing Motion and Content for Video Generation