Drain MPI queues before cleanup

[csegarragonz]

The MpiWorld::destroy() method was working alright until tested from faasm in faasm/faasm#420. There, I add a barrier before calling destroy(). However, our barrier implementation is not fully blocking (ironic).

The way barrier works is:

1. Recv in rank 0 a message from all other ranks.
2. Broadcast from rank 0 when received messages from all ranks.

Before the broadcast is issued, all non-zero ranks are blocked on a recv, however rank 0 does a send (which does not block) and exits the barrier before the other ranks have actually received the barrier. Thus, if this takes a long time, and we clear all the in memory queues before that, undefined behaviour happens.