[SYCL] Implement sycl_ext_oneapi_if_device

[aelovikov-intel]

Specification PR is at #8917.

[aelovikov-intel]

So that we could start removing __SYCL_DEVICE_ONLY__ from the SYCL headers.

[bader]

Please, include minimal required set of headers: queue, buffer, accessor and <sycl/ext/oneapi/experimental/if_device.hpp>.

[aelovikov-intel]

I strongly disagree. The right process to do that is to update the spec/create extension, allocate resources to the activity, cleanup the headers and create automation/checks to automatically cleanup includes and only after that can we start requesting such changes in arbitrary PRs. Until then we have to treat end-to-end tests same way as examples and write them how customers would write their code according to the specification.

[aelovikov-intel]

Failures:

# Intel:
Failed Tests (1):
SYCL :: ESIMD/accessor_local.cpp

# HIP:
Failed Tests (4):
  SYCL :: USM/memops2d/copy2d_device_to_host.cpp
  SYCL :: USM/memops2d/copy2d_host_to_device.cpp
  SYCL :: USM/memops2d/memcpy2d_device_to_host.cpp
  SYCL :: USM/memops2d/memcpy2d_host_to_device.cpp

# CUDA:
Unexpectedly Passed Tests (1):
  SYCL :: DeviceArchitecture/device_architecture_on_host.cpp

covered by

#10138
#10157

and the last is mentioned/awaiting followup in #10173 (comment).

[bader]

@aelovikov-intel, this seems to be ready to merge. What is holding this back?

[gmlueck]

Actually, this should not be merged yet. There was a recent change in the API, and the fn parameter is now passed via perfect forwarding (i.e. as T &&Fn). See #8917.

[bader]

Actually, this should not be merged yet.

Converted to draft. Please, convert back when it's ready to go forward.