Accept NumPy arrays in advanced indexing

[ndgrigorian]

This PR proposes allowing NumPy arrays when indexing usm_ndarray

This can be useful where attempting to use usm_ndarray in a library which generates indices with NumPy, and which expects to only handle devices implicitly through tensor operations (i.e., a library which can take an arbitrary tensor input, but generates indices with numpy.random and expects these indices to work on the input tensor)

Both boolean and integer advanced indices can accept NumPy arrays with these changes

Closes #2053

• Have you provided a meaningful PR description?
• Have you added a test, reproducer or referred to an issue with a reproducer?
• Have you tested your changes locally for CPU and GPU devices?
• Have you made sure that new changes do not introduce compiler warnings?
• Have you checked performance impact of proposed changes?
• Have you added documentation for your changes, if necessary?
• Have you added your changes to the changelog?
• If this PR is a work in progress, are you opening the PR as a draft?

[github-actions]

View rendered docs @ https://intelpython.github.io/dpctl/pulls/2128/index.html

[github-actions]

Array API standard conformance tests for dpctl=0.21.0dev0=py310h93fe807_79 ran successfully.
Passed: 1115
Failed: 6
Skipped: 119

[coveralls]

coverage: 85.886% (+0.004%) from 85.882%
when pulling b90fc90 on allow-numpy-arrays-indexing
into 0dbd495 on master.

[github-actions]

Array API standard conformance tests for dpctl=0.21.0dev0=py310h93fe807_81 ran successfully.
Passed: 1115
Failed: 6
Skipped: 119

[antonwolfy]

I found a suspicious behavior:

q1 = dpctl.SyclQueue(property="in_order")
q2 = dpctl.SyclQueue(property="enable_profiling")

a = dpt.ones(10, sycl_queue=q1)
mask = dpt.zeros(10, dtype='?', sycl_queue=q2)

# raised error as expected:
a[mask]
---------------------------------------------------------------------------
ExecutionPlacementError                   Traceback (most recent call last)
Cell In[30], line 1
----> 1 a[mask]

File dpctl/tensor/_usmarray.pyx:999, in dpctl.tensor._usmarray.usm_ndarray.__getitem__()

File /localdisk/work/antonvol/soft/miniforge3/envs/dpnp_dev/lib/python3.12/site-packages/dpctl/tensor/_copy_utils.py:767, in _extract_impl(ary, ary_mask, axis)
    763     exec_q = dpctl.utils.get_execution_queue(
    764         (ary.sycl_queue, ary_mask.sycl_queue)
    765     )
    766     if exec_q is None:
--> 767         raise dpctl.utils.ExecutionPlacementError(
    768             "arrays have different associated queues. "
    769             "Use `y.to_device(x.device)` to migrate."
    770         )
    771 elif isinstance(ary_mask, np.ndarray):
    772     dst_usm_type = ary.usm_type

ExecutionPlacementError: arrays have different associated queues. Use `y.to_device(x.device)` to migrate.

# no error for 0-d array:
mask = dpt.asarray(1, dtype='?', sycl_queue=q2)
a[mask]
# Out: usm_ndarray([[1., 1., 1., 1., 1., 1., 1., 1., 1., 1.]])

There will be no _extract_impl called, but seems the check is missing in another implementation branch inside __getitem__.

[ndgrigorian]

this is actually an expected behavior: when a zero-dim array is passed as mask (or index), it is brought to the host as a scalar. This especially necessary in the case of boolean masks

see: #1136

[antonwolfy]

Is there missing coercion of usm_type for an input value?

        else:
            exec_q = dpctl.utils.get_execution_queue((exec_q, vals.sycl_queue))
            vals_usm_type = dpctl.utils.get_coerced_usm_type((vals_usm_type, vals.usm_type))

[antonwolfy]

Would it be better to use

Suggested change

	assert dpt.all(dpt.asarray(ii, sycl_queue=q) == y)
	assert dpt.all(x[1:3] == y)

since currently it works only because there is dpt.arange(10, ...) used to allocate data