ENH: Generation of compatable device and dtype pairs for test_from_dlpack

[prady0t]

Closes #440

This is a fairly minimal fix. The new hh.device_dtype_pairs strategy generates pairs of all possible (dtype, device) such that dtype is compatible with device.

Running:

ARRAY_API_TESTS_MODULE=array_api_strict pytest array_api_tests/test_dlpack.py -sv --max-examples=10_000 

succeeds, which previously failed for dtype = float64 on device = no_float64

[ev-br]

Thanks, this seems to work with pytorch now. A couple of further requests below + please remove the now-unused function.

[ev-br]

I'd want to test this on a CUDA-enabled machine locally before merging. The machine is having some issues ATM, meaning I won't be pressing the green button today.

[ev-br]

Got help fixing the CUDA issue, and tested on a multi-GPU system with torch as the backend. The PR seems to work fine. So the only thing which blocks the merge is removing the function which is not used anymore:
https://github.com/data-apis/array-api-tests/pull/443/changes#diff-d40faf59941317167780108980afd61029a89a3c6201c1eaa78e31fb8716ca1aL24-L46

EDIT: the function is test_dlpack.py::_compatible_devices

[prady0t]

Why do we not have CUDA tests in CI? Also, why aren't we testing with different backends in CI, like torch?

[ev-br]

Why do we not have CUDA tests in CI?

These are hard and $$$ to set up. There is no free CI provider, and those which are not free are not easy to set up and maintain either. cf the current SciPy churn.

Also, why aren't we testing with different backends in CI, like torch?

This question seems to pop up for every new person who looks at this test suite :-).
Here's one previous discussion: #298

IME, the most common need is to work in parallel on a branch of array_api_compat and a branch of array-api-tests. This use case is not well served by a CI setup, so what I typically do is $ pip install -e . and $ pushd/popd.

[ev-br]

LGTM now, thank you @prady0t