[SYCL] avoid lock and wait in KernelProgramCache::getOrBuild

[lslusarczyk]

Optimized common path when kernel is already build and exists in the cache. Use BuildStatus already provided by compare_exchange_strong instead of taking lock, calling wait and reading it again.

performance impact

Visible in most benchmarks.
Examples

Instructions decreased from 159.8k to 158.2k over UR baseline 133.7k, that is by 6.1%, see rightmost dots at:

Instructions decreased from 133.4k to 132.3k over UR baseline 119.2k, that is by 7.7%, see rightmost dots at:

And some example of time over L0 (but note time has high variance, so it is less certain)
Time overhead over L0 before: 18.0% (SYCL 17.65us, L0 14.96us), after 15.7% (SYCL 17.13us, L0 14.80us), reduced by ~2.2%

[steffenlarsen]

LGTM!

[lslusarczyk]

@intel/llvm-gatekeepers please merge,
Failed CI also fails on upstream

• E2E (Intel / Ponte Vecchio GPU) fails with HostInteropTask/host-task-failure.cpp timeout, seen also in other PR, e.g. here: https://github.com/intel/llvm/actions/runs/19766696194/job/56642904239?pr=20782
• "SYCL Pre Commit on Linux / build_run_native_cpu_e2e_tests" fail with the same reason (SYCL :: DeviceLib/assert.cpp) as in other PRs, e.g. here https://github.com/intel/llvm/actions/runs/19757562868/job/56612822758?pr=20777