Revisiting Token Pruning for Object Detection and Instance Segmentation