Tear-down parsec_ce after high-level communication

[devreal]

The finalization order is important: parsec_remote_dep_fini finalizes the high-level remote dependency infrastructure (e.g., progress thread), remote_dep_ce_fini tears-down data structures on top of the comm engine (queues). Only then can we finalize the comm engine itself.

Also fixes some weird casts that lead to strange output.

[abouteiller]

Can you document in the PR description what problem it solves

[devreal]

@abouteiller Done 👍

[bosilca]

@devreal I confirm this is the correct order. parsec_remote_dep_fini must be first because it is the blocking function that signal the communication engine to turn off. remote_dep_ce_fini can be next as it unregister the AM for activate and get_data and release the memory management constructs, so it needs the communication engine off.