Move comm profiling initialization into comm thread

[devreal]

The comm thread must call remote_dep_mpi_profiling_init to make sure it's thread-local variable is properly set.

Supersedes #622

[therault]

Remove #ifdef / #endif around pre-declaration

[abouteiller]

the CI errors can be ignored @therault are you satisfied with the changes? If so please merge

[bosilca]

remote_dep_mpi_profiling_init function does two things: registers the dictionary entries for the communication thread and allocate the first bucket for the profiling stream. As far as I recall the dictionary must be created early in the process, this is not something that one can be part of the lazy initialization.

The correct approach would be to split the current remote_dep_mpi_profiling_init function in two, one for registration (that must be called where remote_dep_mpi_profiling_init is called today), and one doing memory allocation (basically the last 2 lines ofremote_dep_mpi_profiling_init), that should be called from the communication thread.

[devreal]

@bosilca please recheck