dpctl should better query SYCL run-times for actually available backends.

[oleksandr-pavlyk]

On a machine without CUDA card, seeting SYCL_BE=PI_CUDA is equivalent to telling SYCL to only expose default host backend.

SYCL_BE=PI_CUDA python -c "import dpctl; dpctl.dump()" shows 4 platforms available, and yet attempting to set it throws errors whose handling is to be improved:

(idp) [16:48:01 linbox work]$ SYCL_BE=PI_CUDA python -c "import dpctl; dpctl.set_default_queue('opencl', 'cpu', 0)"
active queue vector is corrupted.
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "dpctl/sycl_core.pyx", line 721, in dpctl._sycl_core._SyclRTManager.set_default_queue
AttributeError: 'dpctl._sycl_core._SyclRTManager' object has no attribute '_raise_queue_creation_error'

@diptorupd @PokhodenkoSA

[diptorupd]

@oleksandr-pavlyk Is this issue getting fixed by any of the changes we have recently made to device manager and queue manager?

[oleksandr-pavlyk]

set_default_queue is now set_global_queue, and even this could change (can be removed in not so distance future), but it now works correctly:

(idp) [18:33:29 nuc04 dpctl]$ SYCL_DEVICE_FILTER=cuda python -c "import dpctl; dpctl.set_global_queue('opencl:cpu:0')"
Could not find a device that matches the specified filter(s)! -1 (CL_DEVICE_NOT_FOUND)
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "dpctl/_sycl_queue_manager.pyx", line 157, in dpctl._sycl_queue_manager._SyclQueueManager.set_global_queue
    q_obj = SyclQueue(arg)
  File "dpctl/_sycl_queue.pyx", line 233, in dpctl._sycl_queue.SyclQueue.__cinit__
    raise SyclQueueCreationError(
dpctl._sycl_queue.SyclQueueCreationError: SYCL Device 'opencl:cpu:0' could not be created.