Improve static typing for `torch.Size`

[randolf-scholz]

🚀 The feature, motivation and pitch

There are several issues with the current type hints for torch.Size:

pytorch/torch/_C/__init__.pyi.in

Lines 172 to 180 in 9f94710

	# Defined in torch/csrc/Size.cpp
	class Size(Tuple[_int, ...]):
	# TODO: __reduce__
	@overload # type: ignore[override]
	def __getitem__(self: Size, key: _int) -> _int: ...
	@overload
	def __getitem__(self: Size, key: slice) -> Size: ...
	def numel(self: Size) -> _int: ...

It simply subclasses tuple[int, ...], but this causes several typing errors, for example:

import torch
x = torch.Size([1,2,3])
y = torch.Size([4,5,6])
reveal_type(x+y)  # tuple[int, ...], not Size !!!

This is because tuple.__add__ is annotated to return tuple.

Alternatives

No response

Additional context

No response

cc @ezyang @malfet @xuzhao9 @gramster