GPU support

[EnricoDeg]

First steps:

• Nvidia GPUs support by writing a CUDA version of kernels packing and unpacking
• Add examples

[EnricoDeg]

A CUDA backend is currently implemented with the same features of the CPU backend.
Examples use OpenACC but examples should be added using CUDA.
Additional examples using OpenMP offloading might also be of interest.

[EnricoDeg]

Support for AMD GPUs could be easily added because HIP runtime is very comparable to CUDA runtime.
The development can also be done on Nvidia GPUs because HIP supports CUDA in the backend.
The main issue right now is to get access to a feasible environment for development. Options are:

• AMD GPUs, where HIP is usually well supported in terms of software stack
• Nvidia GPUs with HIP software stack installed and well supported