The packing for sending and unpacking from receiving is done with a
kernel that implements a general reordering according to a map.

The implementation again supports both normal MPI and CUDA-aware MPI.