[llvm] [offload] Serialize queue access around launch and sync (PR #198456)

[Akimasa Watanuki via llvm-commits]

Men-cotton wrote:

I got `Concurrent managed access: 0`.

So yes, this environment does not support concurrent managed access. NVIDIA's CUDA on WSL documentation also states that concurrent CPU/GPU access is not supported on WSL 2:
  https://docs.nvidia.com/cuda/wsl-user-guide/index.html#known-limitations-for-linux-cuda-applications

Given that, the local `olLaunchKernelFooTest.SuccessThreaded` failure seems to be an environment/device managed-memory limitation, so I will keep this PR withdrawn.

Thanks for pointing me to the right attribute to check.


https://github.com/llvm/llvm-project/pull/198456