Calling multiple GPU devices from a single thread

[jatkinson1000]

In the example @jwallwork23 wrote in #96 we have n MPI processes for n GPUs.

In a discussion I just had the question was raised - what if we gather data into one process to send to multiple GPUs, or if we have multiple nets and want to cast each one to a different GPU.

My gut feeling is that since Fortran is sequential/imperative, calling multiple nets on different devices from a single process will be no faster than calling them sequentially on the same device (possibly even worse depending on proximity on the node of the CPU to the GPU).

More interesting might be how to assign device indices if we want to gather m >> n processes into n for distribution to n GPUs, but this isn't particularly an "us" problem, nor is it unique to FTorch but occurs more generally in MPI programs.