Add support of usm_ndarray constructor with default dtype on Iris XE

[antonwolfy]

Currently dpctl creates usm_ndarray with default dtype="|f8" which cause an issue on gpu device without double type support.

The below example code cause an exception for Iris XE:

> import dpctl.tensor as dpt
> dpt.usm_ndarray((10,))

ValueError                                Traceback (most recent call last)
Cell In[3], line 1
----> 1 dpt.usm_ndarray((10,))

File dpctl/tensor/_usmarray.pyx:264, in dpctl.tensor._usmarray.usm_ndarray.__cinit__()

ValueError: Device Intel(R) Graphics [0x9a49] does not support |f8 natively.

[oleksandr-pavlyk]

This has been fixed:

In [1]: import dpctl, dpctl.tensor as dpt

In [2]: dpt.usm_ndarray((10,))
Out[2]: usm_ndarray([0., 0., 0., 0., 0., 0., 0., 0., 0., 0.], dtype=float32)

In [3]: import dpctl.memory as dpm

In [4]: dpt.usm_ndarray((10,), buffer=dpm.MemoryUSMShared(80))
Out[4]: usm_ndarray([nan, nan,  0.,  0.,  0.,  0.,  0.,  0., nan, nan], dtype=float32)

[oleksandr-pavlyk]

Please retest and close.

[antonwolfy]

I can confirm the issue is resolved:

import dpctl, dpctl.tensor as dpt

dpt.usm_ndarray((10,))
# Out: usm_ndarray([0., 0., 0., 0., 0., 0., 0., 0., 0., 0.], dtype=float32)

dpctl.__version__
# Out: '0.17.0dev0+232.gd8b9cceed7'