fix(torch-stubs): return torch.return_types named tuples

[yeetypete]

Summary

Add a stub for torch.return_types and use it for the operations that return named tuples. Currently sort, max(dim=...), topk, aminmax, slogdet, etc. return types are annotated as plain tuple[Tensor, Tensor] and .values, .indices, .sign, and the other documented fields are reported as missing-attribute errors.

PyTorch adds these annotations by generating its own return_types.pyi from the output names in native_functions.yaml:

https://github.com/pytorch/pytorch/blob/d6c03540dc0f040e1d8a920d19d9333de6513545/torchgen/api/python.py#L1012-L1033

For simplicity we just write the stubs manually here instead of generating them.

Note: linalg_slogdet is included as well although its missing in the PyTorch generated stubs.

Test Plan

Run test.py.

[stroxler]

Thanks @yeetypete!

Something seems stuck here - auto-import actually failed, and I can't seem to import manually either. Maybe a rebase would help, but another option might be to just close this and open a fresh PR. What ought to happen is that a diff number should usually appear a few minutes after a PR is submitted

[meta-codesync]

This pull request has been imported. If you are a Meta employee, you can view this in D119814478. (Because this pull request was imported automatically, there will not be any future comments.)

[yeetypete]

Apologies for the delay. Rebased and looks like the import worked now :)