NMS with CUDA only

[grimoire]

This PR add a cuda kernel for nms to avoid computation on cpu.

I am not sure if I should call this an "optimization" cause on small input data, cpu performance might even better than gpu. And the data distribution will also affect the performance.

envs:

• Device: RTX 2070super
• CUDA: 11.3
• nvidia driver: 465.19.01
• CPU: Intel(R) Core(TM) i7-9700KF CPU @ 3.60GHz

Test data comes from faster rcnn and demo image. Both nms in rpn and nms in bbox head are tested.

-	old	new
rpn data(4741 boxes)	0.98435ms	0.88798ms
rcnn data(583 boxes)	0.23121ms	0.22827ms
random data(500 boxes)	0.23056ms	0.29762ms
random data(1000 boxes)	0.26279ms	0.41685ms
random data(5000 boxes)	1.27405ms	1.34129ms
random data(10000 boxes)	3.62377ms	2.74194ms
random data(20000 boxes)	21.07619ms	4.80819ms

Real data might have clustered bboxes, which reduce gpu computations. I guess that is why real data performance is better than random data.