Inconsistency between index_select and __get_item__

[smorad]

🐛 Describe the bug

We should be able to index using torch.int32 using the bracket notation. It is possible using index_select:

>>> idx = torch.tensor([0]).int()
>>> a = torch.rand(3)
>>> a[idx]
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
IndexError: tensors used as indices must be long, byte or bool tensors
>>> a.index_select(0, idx)
tensor([0.6999])

When dealing with large sets of indices (e.g. graph edge lists), this means we have to use more memory.

Versions

master

[vadimkantorov]

Related: #61819, pyg-team/pytorch_geometric#5087

It would be nice to support other integer dtypes, and without upcasts using more memory (maybe via jiterator or some jit/codegen?)