Unsupervised Learning of Shape and Pose with Differentiable Point Clouds