On the Adversarial Robustness of Camera-based 3D Object Detection