Make MPI optional

To help lower the barrier to entry, we want to make MPI an optional feature. This will require splitting modules that use MPI calls into single process components and multi-process components. If MPI is detected, then we ought to build MPI enabled libraries in addition to the single process libs.

This will ultimately necessitate breaking up the library artifacts that are created into a few groups

* selfcore
* self_[cpu-arch]
* self_[cpu-arch]_mpi
* self_[cpu-arch]_[gpu-arch]
* self_[cpu-arch]_[gpu-arch]_mpi

When users build applications with self, the intention would be to link against `-lselfcore` and the appropriate library for the architecture they want to build for.