Implement torch.diagonal for named tensors (#30193)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/30193
Featuring:
- Added a NoNamesGuard::reset() function that sets NamesMode back to
what it was before the guard. This makes it so that we don't have to
create a new context to run code in an unnamed way.
- Added a diagonal(Tensor, *, Dimname outdim, Dimname dim1, Dimname dim2, int64_t offset=0)
overload. All of the non-tensor arguments are keyword only for
readability purposes; something like `tensor.diagonal("A", "B", "C")`
would be really confusing.
Test Plan: - Added new tests
Differential Revision: D18638363
Pulled By: zou3519
fbshipit-source-id: ea37b52a19535f84a69be38e95e569e88f307381
