mmtracking with torch >= 2.2.2 fails with error: RuntimeError: nms_impl: implementation for device cuda:0 not found.

[nvijayrania]

Checklist

1. I have searched related issues but cannot get the expected help.
2. The bug has not been fixed in the latest version.

Describe the bug
There seems to be a circular dependency when I'm using mmtracking with torch 2.2.2 and higher versions.

• With mmcv-full 1.7.2 & mmtracking 0.14.0, the finetuning run fails with RuntimeError: nms_impl: implementation for device cuda:0 not found.
• With mmcv 2.2.2 & mmtracking 0.14.0, there are multiple import failures and assertion errors such as AssertionError: MMCV==2.1.0 is used but incompatible. Please install mmcv>=1.3.17, <2.0.0.

Bug fix
Could we please remove this assertion and make mmtracking compatible with latest mmcv 2.x versions to allow torch version >= 2.x.