Relax array dimension rules on array MPI communi
Right now, we assume that multi array merged and dispatched other the communicator should have the same dimension as the grid.
This is not the case, just nr needs to comply. ANd it make CUDA testing and debuging complicated.