Add experimental HIP support

[G-071]

This PR adds the first experimental support for running Octo-Tiger on AMD GPUs.

While the effort regarding a (performance-portable) AMD GPU port is still ongoing, I think this is a good point to get some of the changes into master (before the PR gets too unwieldy).

Changelog:

• Added HIP kernel variants for most ¹ major CUDA kernels (basically by hipifying the API)
• Reworked CUDA/HIP Kernels for increased performance (separated Stencils into more blocks)
• Added support for running most ¹ Kokkos kernels on the HIP backend via the hpx-kokkos hip executors
• Added CLI options to actually use the new kernels (device kernel types: HIP, KOKKOS_HIP)
• Added tests for all new kernels
• Added Jenkins Pipeline for AMD GPU tests on the Rostam MI100 nodes
• Added Jenkins Pipeline for the Stuttgart Radeon VII Pro node
• Added warnings about HIP support still being experimental...

Note:

• Still does not quite achieve the same performance as we get on NVIDIA devices -> needs more work!
• CLI parameter names need to be renamed ("cuda_streams_per_gpu" is an odd name for the number of Kokkos / HIP executors..) - as this will change the interface for all existing scripts, I would postpone this task for a later PR

Build instructions:

• Set the cmake variable OCTOTIGER_WITH_CUDA=OFF and OCTOTIGER_WITH_HIP=ON
• Experimental toolchain branch available here

Remaining ToDos

• Cleanup

[1]: Except the p2m kernel as yet, as it requires more effort to port but only constitutes a small percentage of the runtime

[diehlpk]

I added two comments which should be addressed.

[diehlpk]

LGTM!