Refactor detections algorithms

[mli]

Rewrite the detection part: removed the customized operators and reimplemented them with numpy, add faster-rcnn training.

We needs the following steps before merging this PR

1. @reminisce @haojin2 the current implementation runs on @haojin2 's np_boolean_mask_assign_broadcast brunch, but not on the nightly build. It also has several FIXMEs in the codes. Please make sure there is a nightly build runs the codes with FIXMEs fixed (except for use_numpy in NMS as our implementation is too slow).
2. both faster-rcnn and ssd accuracies aren't great. @zhreshold can you please have a look if there is any bug or just the hyper-parameters aren't correct after the previous step is finished.
3. need to polish the writing
4. (optional) add yolo-v3 implementation

By submitting this pull request, I confirm that you can use, modify,
copy, and redistribute this contribution, under the terms of your
choice.

[haojin2]

apache/mxnet#17131 has been merged, this PR can depend on latest nightly starting from tomorrow.

[zhreshold]

@haojin2 The latest nightly mxnet is mxnet_cu100mkl-1.6.0b20200215-py2.py3-none-manylinux1_x86_64.whl which still shows

MXNetError: Traceback (most recent call last):
  File "src/operator/numpy/np_boolean_mask_assign.cc", line 260
MXNetError: Unknown type enum 7

So I think it's not available yet?

[AnirudhDagar]

The following scratch methods are now available through #1565 #1579 #1580:

• box_corner_to_center
• box_center_to_corner
• multibox_prior
• box_iou
• match_anchor_to_bbox
• offset_boxes
• multibox_target
• offset_inverse
• nms
• multibox_detection