Time-to-Label: Temporal Consistency for Self-Supervised Monocular 3D Object Detection