- Rename keyword "gpu" -> "nvidia-gpu"
- Add keyword "amd-gpu"

The GPU logic has been implemented in the OpenMP code paths in ELPA2.
Currently, this implies that _internal_ to ELPA2, the number of OpenMP
threads is set to one (independent of how many threads the calling
application uses) and the original value is restored at the end of ELPA.
Though this is not the general case, this is _not_ a limitation, since
in the GPU case no work is done on the CPU and thus no threading is
helpful

* cudaMallocHost
* cudaFreeHost
* cudaHostRegister
* cudaHostUnregister

* Switch to a simple non-WY algorithm
* Unify real and complex cases
* Update reduction kernel
* Use __shfl_xor_sync for warp reduce (CUDA 9+)
* Support 2^n block size, n = 1,2,...,10
* Use templates when possible
* Clean up unused CUDA functions
* Increase default stripe width when using GPU

with cudaDeviceSynchronize

When profiling the GPU version, NVTX can be used to highlight the
corresponding regions of the code in the timeline of the profiling tool
(nvvp or nsight systems). This is very useful to correlate what happens
on the GPU with what part of the code we are in.

The GPU initialization is actually quite constly, e.g. on Minsky it
takes roughly 0.7s. That is hurting performance for small matrices.
Thus a check has been added and now GPU should be initialized only the
first time.

on misnky