use post() instead dispatch() for async_await()

[olk]

• dispatch() might queue the completion handler or invoke it immediately
• this might result in problems if this code runs in completion
  handlers like coroutines where you must not resume an already running coroutine
• using post() solves this issue

[gulachek]

This might fix another issue you found, but this documented issue (196) isn't fixed by the diff I'm seeing. If you see the AsyncReadStream requirements, the async_pipe needs to implement a get_executor member function.

[olk]

This patch from above seams to solve our production code.

The problem was that async_coroutine::push() was invoked from sigchld_service::async_wait() while the stackful coroutine was already running.

I refer to the issue #195 not pull-request #196 - my mistake, sorry.

[klemens-morgenstern]

I think the outer one should be dispatch, but the inner one should post.

[olk]

Klemens, I'll try your suggestion at our production code...
But could you give me an explanation why you think that the outer one should call dispatch()?

[klemens-morgenstern]

The outer just needs to make sure it's on the right executor. I actually think the fix would be to post these invocations:of h.

                    if (pid_res < 0)
                        h(-1, get_last_error());
                    else if ((pid_res == pid) && (WIFEXITED(status) || WIFSIGNALED(status)))
                        h(status, {}); //successfully exited already