Delving into the Pre-training Paradigm of Monocular 3D Object Detection