ENH: 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
• minor device-related tweaks

[Copilot]

Pull Request Overview

This pull request implements a workaround for torch.asarray’s device propagation issue by adjusting imports and type annotations, along with minor device‐related tweaks across multiple array API compatibility modules.

• Update type hints for device parameters and return types (from str to Device)
• Replace the custom Device import with torch.device in torch/_typing.py
• Introduce a new asarray function in torch/_aliases.py to address dtype propagation

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	Updated import of Device from torch and modified the all order
array_api_compat/torch/_info.py	Adjusted docstrings to reflect Device type for default device methods
array_api_compat/torch/_aliases.py	Added a new asarray function and removed redundant torch.asarray calls
array_api_compat/numpy/_info.py	Changed device type annotations from str to Device
array_api_compat/dask/array/_info.py	Changed device type annotations from str to Device
array_api_compat/cupy/_info.py	Updated device type annotations and added Notes in docstring
array_api_compat/common/_aliases.py	Updated inline comments to include Dask in creation functions

Comments suppressed due to low confidence (3)

array_api_compat/torch/_typing.py:3

• The change to import Device directly from torch (using torch.device) may affect the expected Device behavior compared to the previous custom definition. Please verify that this change preserves the intended device propagation semantics throughout the code.

from torch import device as Device, dtype as DType, Tensor as Array

array_api_compat/torch/_aliases.py:228

• In the new asarray function, only torch.Tensor inputs have their device set automatically. Consider verifying that non-tensor objects are handled correctly when device is None.

if device is None and isinstance(obj, torch.Tensor):

array_api_compat/torch/_aliases.py:308

• Removal of the explicit torch.asarray conversion in functions like prod, sum, any, and all might lead to unintended type issues if the input is not already a tensor. Confirm that callers always supply a tensor or that the new asarray function is applied upstream to avoid runtime errors.

    -    x = torch.asarray(x)

[crusaderky]

CI failures unrelated (#297)

[ev-br]

close/reopen to rerun the CI

[crusaderky]

I can't reopen a PR someone else closed.
Repoened in #299.

[ev-br]

Apologies for the glitch. No idea why it's not got reopened.