Document why combining pipeline<thread_scope_thread>/barrier<thread_scope_thread> with memcpy_async is legal

[gonzalobg]

According to the thread scope documentation:

A thread scope specifies the kind of threads that can synchronize with each other using a primitive such as an atomic or a barrier.

Each thread (CPU or GPU) is related to itself by the thread thread scope, specified with thread_scope_thread.

This example, however, suggests that multiple threads of execution - the current thread and the threads performing the memcpy_async - can be synchronized using the synchronization primitives with thread_scope_thread.

A fix could be to change the documentation to:

The thread thread scope relates each thread (CPU or GPU) and its ??? threads to itself.

where ??? would be a name for the set of threads that includes the ones used by the memcpy_asyncs.

[griwes]

I don't think our current docs imply that there's inter-thread synchronization, because I don't think we mention any side threads in memcpy_async docs; we only say that we issue an async operation on the current thread and that's it.

@ogiroux @c0riolis do you want to comment on this?

[ogiroux]

We need to say that the _async operations are performed as-if in another thread, but there is a special exception where this helper thread is related to the requestor thread via all scope relationship, including surprisingly the _thread scope.

[griwes]

@gonzalobg does NVIDIA/libcudacxx#371 fix this?