Tidy up GPU code, speed up and simplify

[smutch]

I think we can probably simplify the GPU code quite a bit.

• Look at maybe simplifying error handling using the examples in the CUDA developer docs.
• Use streams to improve performance.
• Get CUFFT working with MPI and multi-GPU (hopefully we can do this if we have one rank orchestrate things).
• Allow Mhysa to specify what GPU each rank should connect to, that would allow us to do away with the use of MPS.