Fences the async proxy in CUDA, to make shared memory available to it. Does not implement
Synchronizes, because it works only as a memory availability barrier with an outside chip.
It does not synchronize the actual threads, and is typically called only by the TMA leader.
Scope that the synchronization should apply to. This is a minimum, when fine-grained control
is not available it should synchronize at the smallest scope that includes this scope
(i.e. SyncScope::Plane may be implemented by a workgroupBarrier())