MPI compliant custom reductions

[patflick]

See here: open-mpi/ompi#1462

Two options really:

1. Use MPI_Pack/Unpack for MPI compliant copying of the in/inout buffers inside the custom reduction operation.
2. Or use MPI_Type_get_envelope and manually copy over each member as they appear in the MPI datatype. I could also cache the envelope information myself inside the mxx::datatype and provide a MPI_Datatype safe copy operation in that class (e.g. copy member by member).