[ndgrigorian]

This PR makes binary elementwise functions with a Python or NumPy scalar operand asynchronous, closes #3148

A scalar operand (2 * x, x == 0, x * np.float32(2)), the scalar of clip and where, and a scalar assigned through basic indexing (x[0] = 1) were turned into a 0-d array by _copy_numpy_ndarray_into_usm_ndarray. That copy waits for all work already submitted to the queue, so each of these calls blocked the host until the device was idle. _copy_from_numpy_into and dpnp.tensor.asarray now fill 0-d input with _full_usm_ndarray, ordered after the queue's submitted events, instead. Arrays with ndim > 0 still go through the copy

The value is cast on the host the way the copy kernel would cast it, so results don't change

• Through float32/complex64 on devices without fp64
• Complex to real keeps the real part
• Float to integer goes through int64, so it wraps like the copy kernel where a direct NumPy cast saturates (e.g. 1e10 into int16 gives -7168)
• No NumPy cast warnings

The new tests patch _copy_numpy_ndarray_into_usm_ndarray to raise, and cover asarray of Python scalars, NumPy scalars and 0-d arrays of every dtype, casts to another dtype, fp64 scalars on devices without fp64, indexed assignment, and binary operators, out= and in-place operators with a scalar operand. They fail without the fix (131 cases). The affected dpnp.tensor tests and dpnp/tests/test_arraycreation.py pass on a Level Zero iGPU and the OpenCL CPU device. The fp64-less test only skipped here, since both devices support fp64

With the reproducer from #3148 on the iGPU, the host time of one call issued right after a ~75 ms matmul:

call	master	this PR
x * x	0.45 ms	0.51 ms
2 * x	69.37 ms	1.23 ms
x == 0	80.70 ms	0.92 ms
x * np.float32(2)	75.50 ms	1.32 ms
x[0] = 1	76.07 ms	0.43 ms

• 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?

[github-actions]

View rendered docs @ https://intelpython.github.io/dpnp/pull/3182/index.html

[github-actions]

Array API standard conformance tests for dpnp=0.22.0dev1=np2py314h8d9cdd5_7 ran successfully.
Passed: 1491
Failed: 0
Skipped: 6

[coveralls]

coverage: 84.397%. remained the same — make-scalar-operands-in-binary-ops-asynchronous into master