ParallelMergeMesh merge() hangs for MPICH 3.3 for some cases.

[Michel de Messieres]

When we switched to MPICH 3.3 from 3.1.4 we observed Moab would hang in some instances. This is happening in the ParallelMergeMesh merge() call when it calls ghost exchange.

Attached is a simplified code which will run on MPICH 3.1.4 and hang for 3.3.

This is a 2D 16x64 mesh though many other combinations run fine on MPICH 3.3. So it seems specific to a small subset of dimensions.

In the test a structured database is created and then the verts and cells are copied to a new unstructured database.

Tested on the most recent release of master.