Add hipSYCL curand support

[nilsfriess]

Description

This PR adds support to use the cuRAND backend with hipSYCL.
The approach is similar though simpler than #144: Since hipSYCL does not support host_task but instead implements the extension hipSYCL_enqueue_custom_operation, a wrapper function onemkl_curand_host_task is introduced that calls either host_task or hipSYCL_enqueue_custom_operation depending on which compiler is used.

Further, calls to curandSetStream before calling the cuRAND generate-functions were added. This is important since previously, the call to wait_and_throw after submitting the CUDA calls might not wait for the cuRAND random number generation to finish if it was started on a different stream. In fact, increasing the amount of random numbers generated in the unit tests (and thereby increasing the time spent generating numbers) made some tests fail (both when using hipSYCL and dpc++), since the result might be read before random number generation is finished. Now all tests pass also when increasing the number of random samples to be generated.

Test logs

The tests were executed on Ubuntu 20.04 using CUDA 11.6.
curand_dpcpp_test.log
curand_hipsycl_test.log

Checklist

• Do all unit tests pass locally? Attach a log.
• Have you formatted the code using clang-format?

[aelizaro]

Hi @nilsfriess, thank you for the PR, looks good to me! @mkrainiuk, could you, please, also take a look?

[mkrainiuk]

Thank you for the PR, it also looks good to me. Does it make sense to update also our documentation with RNG supported compilers?

[nilsfriess]

I think there are multiple places where the documentation is not quite up to date with regard to hipSYCL support (e.g. here and here it is not mentioned that hipSYCL can also be used for the BLAS backend with CUDA).
Should I make an extra PR for the documentation updates or just include it in this one?

[mkrainiuk]

I think there are multiple places where the documentation is not quite up to date with regard to hipSYCL support (e.g. here and here it is not mentioned that hipSYCL can also be used for the BLAS backend with CUDA). Should I make an extra PR for the documentation updates or just include it in this one?

Good catch, we need to update them too. I'm fine with both options, if you prefer extra PR we can merge this PR as is.

[nilsfriess]

There weren't too many places where the documentation was not up to date, so I just added the changes to this PR.

[mkrainiuk]

There weren't too many places where the documentation was not up to date, so I just added the changes to this PR.

Thank you! Documentation looks good to me. Please let me know if there anything else you want to add or we can merge it now.

[nilsfriess]

Thanks, you can merge it :)