SCIZOR: A Self-Supervised Approach to Data Curation for Large-Scale Imitation Learning