work around MPI implementations that are not count-safe

[jeffhammond]

We saw in #9 that implementations that have count-safety issues are a problem. Such issues can be avoided in many cases, although MPI_THREAD_MULTIPLE support and MPI wildcards may be a problem, since splitting messages may lead to false matching in these instances.

We should implement the multiple-messages implementation currently available only for blocking reductions everywhere possible. One way to do it for all the collectives is to use the P2P back-end and then have a P2P version that breaks up large transfers. However, this is problematic if cross-matching can occur, so we should create a duplicate communication (on-the-fly to avoid having an init function, even though it is potentially unsafe).