Would be useful to have an example that uses mpi4py with Warp arrays using the CUDA array interface. Can make a Python version of https://github.com/NVIDIA/multi-gpu-programming-models/tree/master/mpi_overlap