I think there's a descriptor leak in Unix.close_process_full and
possibly other close_process* functions. If there's an error in
closing one of the descriptors, I think the correct behavior would be
to close the rest nevertheless before throwing a Unix_error exception.

I think this can happen in some conditions (SIGPIPE?) even without
having explicitly closed some of the descriptors. Now I have a
situation where a mere Unix.close_process_full will result in a
descriptor leak but explicitly trying to close them one by one will
not.

	Sami