Fix derivatives of norm(p=inf)
Following up on https://github.com/pytorch/pytorch/pull/51099#discussion_r583323915, we fix these derivatives, as they were incorrect until now.
As described in the note, the better solution would be to use vectorised operations on the preprocessing operation when reducing on CPU. It's not clear how difficult that may be.
Fixes https://github.com/pytorch/pytorch/issues/67517
Pull Request resolved: https://github.com/pytorch/pytorch/pull/78105
Approved by: https://github.com/ngimel
