Truncate-Split-Contrast: A Framework for Learning from Mislabeled Videos