CaSPR: Learning Canonical Spatiotemporal Point Cloud Representations