[Relay] Support dcnv2(mask) for pytorch in tvm

[Fvoiretryzig]

Add dcnV2 support according to this repository

[zhiqwang]

I have a question there, TorchVision has also implemented a version of DCNv2： https://github.com/pytorch/vision/blob/main/torchvision/csrc/ops/deform_conv2d.h , it should support exporting to torchscript. Could we also test this version?

[masahi]

I'm aware that our deform covn2d support lacks mask support. If that is what this PR is about, this is great!

Please clean up your change (I see unrelated removal of existing code)

[Fvoiretryzig]

I have a question there, TorchVision has also implemented a version of DCNv2： https://github.com/pytorch/vision/blob/main/torchvision/csrc/ops/deform_conv2d.h , it should support exporting to torchscript. Could we also test this version?

I wrote this operator support for compiling CenterTrack Model in tvm. It uses dcnv2 from this repository. This implementation doesn't support dcnv2 in torchvision, but it can modify the tvm pytorch frontend part to support. Does this need to change to support torchvision? @zhiqwang

[Fvoiretryzig]

I'm aware that our deform covn2d support lacks mask support. If that is what this PR is about, this is great!

Please clean up your change (I see unrelated removal of existing code)

I have updated this repository. @masahi @tqchen

[JiaqingFu]

can this pr merged？ @Fvoiretryzig @masahi @tqchen