BUG: torch.asarray device propagation

[crusaderky]

• Workaround to asarray: device does not propagate from input to output after set_default_device pytorch/pytorch#150199
• Also fixes device propagation in prod, sum, any and all, which had the same issue induced by array-api-compat

[Copilot]

Pull Request Overview

This PR implements a workaround for a known PyTorch issue with torch.asarray device propagation and improves device handling across several functions. The changes include:

• Updating imports and all ordering in torch/_typing.py to source the Device type directly from torch.
• Revising type annotations and docstrings across torch, numpy, dask, and cupy _info.py files to reflect Device rather than a string.
• Introducing an asarray helper in torch/_aliases.py and removing redundant explicit conversion calls in prod, sum, any, and all.

Reviewed Changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
array_api_compat/torch/_typing.py	Adjusted import order and all for device type propagation.
array_api_compat/torch/_info.py	Updated docstrings for default_device and default_dtypes to use Device.
array_api_compat/torch/_aliases.py	Added asarray helper and removed explicit torch.asarray conversions in several functions.
array_api_compat/numpy/_info.py	Updated device type annotation in function docstrings.
array_api_compat/dask/array/_info.py	Updated device type annotation in function docstrings.
array_api_compat/cupy/_info.py	Updated device type annotation and added Notes for default_device.
array_api_compat/common/_aliases.py	Updated a comment on creation functions to reflect support for Dask.

Comments suppressed due to low confidence (4)

array_api_compat/torch/_aliases.py:283

• In function 'prod', the explicit conversion to a tensor was removed. Ensure that 'x' is already a torch.Tensor or add conversion logic to prevent potential runtime errors.

-    x = torch.asarray(x)

array_api_compat/torch/_aliases.py:319

• In function 'sum', the removal of the torch.asarray conversion may lead to errors if 'x' is not already a tensor. Consider validating or converting 'x' as needed.

-    x = torch.asarray(x)

array_api_compat/torch/_aliases.py:349

• In function 'any', the removal of the torch.asarray conversion assumes that 'x' is already a tensor. Verify that this holds for all call sites to prevent possible runtime issues.

-    x = torch.asarray(x)

array_api_compat/torch/_aliases.py:374

• In function 'all', ensure that removing the explicit torch.asarray conversion will not cause issues when 'x' is not a tensor. It may be necessary to handle non-tensor inputs safely.

-    x = torch.asarray(x)

[lucascolley]

LGTM, thanks!