Add allow_negative_crop in RandomCrop

[Johnson-Wang]

1. Add allow_negative_crop in RandomCrop
2. Resolve a TODO: check whether we can keep the image regardless of the crop.
3. Find out the reason of a sudden boost of nas-fpn performance using the RandomCrop in Fixed bugs in transform.py #2810. In Fixed bugs in transform.py #2810, the allow_negative_crop was implicitly set to True when there is gt_bboxes_ignore flag and gt_bboxes_ignore is an empty tensor.
   I will keep gt_bboxes_ignore=False in this version and provide the performance with gt_bboxes_ignore=True in the nas-fpn README.

Conclusion after re-running the experiments

The performance is very close for coco dataset with gt_bbox_ignore=True or gt_bbox_ignore=False. Both of these are higher than that reported in the model zoo

	before #2810	#2810	gt_bbox_ignore=True	gt_bbox_ignore=False
retinanet_fpn_crop640	37.4	37.9	37.8	37.9
retinanet_nas_fpn_crop640	40.2	40.5	40.6	40.7