[llvm] [Offload] Make olLaunchKernel test thread safe (PR #149497)

[Ross Brunton via llvm-commits]

================
@@ -2302,8 +2304,11 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
     // Once the stream is synchronized, return it to stream pool and reset
     // AsyncInfo. This is to make sure the synchronization only works for its
     // own tasks.
-    AsyncInfo.Queue = nullptr;
-    return AMDGPUStreamManager.returnResource(Stream);
+    if (RemoveQueue) {
----------------
RossBrunton wrote:

Liboffload contains this:
```c++
Error olWaitQueue_impl(ol_queue_handle_t Queue) {
  // Host plugin doesn't have a queue set so it's not safe to call synchronize
  // on it, but we have nothing to synchronize in that situation anyway.
  if (Queue->AsyncInfo->Queue) {
    if (auto Err = Queue->Device->Device->synchronize(Queue->AsyncInfo, false))
      return Err;
  }


  // Recreate the stream resource so the queue can be reused
  // TODO: Would be easier for the synchronization to (optionally) not release
  // it to begin with.
  if (auto Res = Queue->Device->Device->initAsyncInfo(&Queue->AsyncInfo))
    return Res;

  return Error::success();
}
```

This has to be done atomically so that, for example, we don't try to synchronise an absent queue. I could add a mutex to `ol_queue_impl_t`, but I figured it'd be better to just implement what the comment says. Specifically, we avoid dropping the AsyncInfo just to immediately recreate it right after.

https://github.com/llvm/llvm-project/pull/149497