CaSPR: LearningCanonicalSpatiotemporal PointCloudRepresentations