[SYCL] Treat profiling as not supported if submit timestamps are not

[sergey-semenov]

With the recent fix to command_submit profiling information query it now requires that the plugin supports piGetDeviceAndHostTimer API. Because of that, we should not report queue profiling as a supported aspect unless both the underlying pi_queue supports profiling and piGetDeviceAndHostTimer is available.

In addition, add a feature_not_supported exception when a queue is constructed with enable_profiling property and a device that does not support profiling.

[sergey-semenov]

Note that the spec does not currently state that the exception must be thrown in the case described in the commit message, but I don't expect it to be a controversial change (KhronosGroup/SYCL-Docs#359)

[sergey-semenov]

Looks like ESIMD tests assume that queue profiling is supported despite that not being the case with ESIMD emulator (it's just actually checked with this patch). I think the tests should be modified so that profiling is optional there.

I noticed, however, that several profiling related functions from the ESIMD emulator backend (piGetDeviceAndHostTime, piEventGetProfilingInfo) return PI_SUCCESS instead of an error despite not being able to do anything meaningful. @dongkyunahn-intel Is that intentional?

[dongkyunahn-intel]

@sergey-semenov , as far as I remember, when I brought up ESIMD_EMULATOR backend, piEventGetProfilingInfo had to return PI_SUCCESS for the backend to pass some tests even if the implementation does nothing meaningful while profiling under emulation is not critical. That was the reason why warning message was added in the function. I was not aware that piGetDeviceAndHostTime was added.

[sergey-semenov]

/verify with intel/llvm-test-suite#1604

[sergey-semenov]

/verify with intel/llvm-test-suite#1604

[sergey-semenov]

@dm-vodopyanov @intel/llvm-reviewers-runtime Could you please take a look?