Check cuda-aware mpi at runtime
If LinkTest is run with --use-gpu-memory, but the MPI is not cuda aware it will segfault at MPI_Send, try to detect this early and give a better error message See https://gist.github.com/K-Wu/6c353273aafe9a4eaaa344a8b74475b6 for examples. However some setups might not be detectable.