Hotfix: profiling initialized from the wrong thread in comm engine

[therault]

After we introduced lazy initialization of the comm engine, we left the profiling initialization for the remote_dep comm engine be called from the main parsec thread. That initialization must be called from the communication thread, this patch moves it to the actual comm thread. (bug found and fixed by @devreal).

[bosilca]

What is the real change here ? This patch moves a bunch of code around for no obvious reason.

[devreal]

Thanks @therault!

@bosilca The real change is that remote_dep_mpi_profiling_init should be called from within remote_dep_dequeue_main (the comm thread's main function) instead of remote_dep_ce_init (called by the main thread to spin up the communication backend). The rest is moving code around for correct ordering.

[bosilca]

Can we just pre-decleare the function instead of moving all the code around ?

[abouteiller]

We will use #626 instead