Fix native_batch_norm_backward returning non-channels_last_3d grad (#107270)
Fix #107199
Checked out https://github.com/pytorch/pytorch/pull/106104 which caught this locally and verified that https://github.com/pytorch/pytorch/blob/551124f67090b7c672e9833ed10f41d31c57747c/torch/testing/_internal/common_modules.py#L2635-L2642 with the `p['device'] == 'cuda'` part shifted to `device_type = 'cuda'` now succeeds
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107270
Approved by: https://github.com/albanD
