Replace NMS GPU kernel in object detection sample with torchvision's implementation

Make sure that inference speed in eval mode and accuracy is not significantly affected.

In case torchvision’s NMS impl is slower, clean up the existing GPU kernel by removing the empty wrapper in nms.cpp

[SKholkin]

There is no such big difference in both inference speed and accuracy Actually PyTorch impl even a little bit faster and more accurate: ssd300 PyTorch impl: Averaged detect for batch: 1.163s Mean AP = 0.7831 Old FB kernels: Averaged detect for batch: 1.192s Mean AP = 0.7828

ssd512 PyTorch impl: Averaged detect for batch: 0.819s Mean AP = 0.8044 Old FB kernels: Averaged detect for batch: 0.843s Mean AP = 0.8026

[ljaljushkin]

Custom NMS CUDA kernel from our extension uses top-k to cut a number of predictions. It’s aligned with implementation from OpenVINO: https://github.com/openvinotoolkit/openvino/blob/master/inference-engine/src/mkldnn_plugin/nodes/detectionoutput.cpp#L568

NMS from Torchvision doesn’t do that, and it seems we can’t cut a number of detections before and after function to make results equivalent to the NMS from our extension because the top-k is used internally on par with a full array of predictions.

int boxes_num = std::min(boxes.size(0), top_k);
scalar_t* boxes_dev = boxes_sorted.data<scalar_t>();
nms_kernel<<<blocks, threads>>>(boxes_num,
                                nms_overlap_thresh,
                                boxes_dev,

Without the top-k parameter, the implementations are equivalent, so I believe it’s OK to keep the custom NMS GPU kernel.

@vanyalzr, please re-open the issue, if you don’t agree or have other ideas to discuss.