HF RL Explorer

Asynchronous FlagCX collectives can incorrectly appear complete when waiting on the returned work handle

Asynchronous FlagCX collectives can incorrectly appear complete when waiting on the returned work handle: a task in LegoFlow-SWE (Harbor dataset). WorkFlagCX::wait() must guarantee that it waits for all GPU work associated with the WorkFlagCX instance to finish. In particular, the wait operation…

The task

`WorkFlagCX::wait()` must guarantee that it waits for all GPU work associated with the `WorkFlagCX` instance to finish. In particular, the wait operation should ensure that the event/callback block associated with the work is actually invoked and that the stream has reached the completion point for the queued…

Part of Lego-X/LegoFlow-SWE.