`conj` should delegate to `torch.conj_physical`

[lucascolley]

See scipy/scipy#20772 (comment)

[asmeurer]

So the issue is that torch.conj works by setting a conjugation bit, which means np.asarray(torch.conj(x)) won't work. torch.conj_physical does an actual conjugation. This is considerably less efficient (torch.conj runs in constant time whereas conj_physical has to compute the conjugation), but it seems that things like np.asarray or even __dlpack__ won't work on tensors with the conjugation bit.

>>> a = torch.linspace(0, 10, 10000) + torch.linspace(0, 10, 10000)*1j
>>> np.asarray(torch.conj(a))
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/aaronmeurer/miniconda3/envs/array-apis/lib/python3.12/site-packages/torch/_tensor.py", line 1087, in __array__
    return self.numpy()
           ^^^^^^^^^^^^
RuntimeError: Can't call numpy() on Tensor that has conjugate bit set. Use tensor.resolve_conj().numpy() instead.
>>> torch.conj(a).__dlpack__()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/aaronmeurer/miniconda3/envs/array-apis/lib/python3.12/site-packages/torch/_tensor.py", line 1480, in __dlpack__
    raise RuntimeError("Can't export tensors with the conjugate bit set")
RuntimeError: Can't export tensors with the conjugate bit set

This seems like something that the array API ought to be able to support. But maybe the issue is that PyTorch should just realize the conjugation automatically when converting to NumPy or DLPack instead of requiring the user to do it.

[rgommers]

But maybe the issue is that PyTorch should just realize the conjugation automatically when converting to NumPy or DLPack instead of requiring the user to do it.

Yes, I think that would be the correct thing to do, rather than PyTorch erroring out.

[asmeurer]

Well the error seems to be intentional, so there might be an explicit design decision there.