multiple-rank-mpi-launch-with-openmpi-is-suboptimal

[egstern]

OpenMPI has a bizarre default way of distributing MPI ranks among processors.
For 2 or fewer MPI ranks, it assigns ranks to cores. For more than 2, it
assigns ranks by socket. This results in extremely long test execution
times for 3 or more MPI ranks on some systems. The solution is to force
OpenMPI to always use map-by core.

[s-sajid-ali]

This was considered against in #8