Add MultiGPU Support

[G-071]

This PR adds experimental multi-GPU support for all CUDA/KOKKOS kernels in Octo-Tiger.

To do so, it mostly uses the additions in STEllAR-GROUP/hpx/pull/6284 and SC-SGS/CPPuddle/pull/22.
Accordingly, this PR uses the new functionality defined in those PRs. To work correctly, it also

• ensures that the required constant data is initialized on all GPUs.
• sets the correct device before all Kokkos calls.
• and, of course, the PR adapts to all API changes, necessary to initialize the executors on all devices.

Overall, it seems to work well on the machines I tested it on. I will probably re-structure the MultiGPU part of the gravity solver a bit in the future, but I think that is a subject for another PR. This one gives us the basic functionality for now!

[G-071]

I reworked this PR (and CPPuddle) as I did not quite like the way the previous MultiGPU support was working! However, it is ready for review now!

New Features/Fixes:

• Add MultiGPU capabilities for gravity and hydro solver (works with CUDA / HIP and Kokkos and NVIDIA and AMD devices).
• Add workaround for the initialization issues on AMD devices
• Add option for separate polling threads
• Add option for additional time-per-timestep analysis
• Add HPX performance counter for kernel launches and work aggregation statistics
• Register new CPPuddle performance counters for additional memory statistics

Interface changes:

Renamed existing parameters (to make clearer that they are not CUDA specific anymore):

• --cuda_number_gpus -> --number_gpus
• --cuda_streams_per_gpu -> --executors_per_gpu
• --max_executor_slices -> --max_kernels_fused
• --cuda_buffer_capacity -> --max_gpu_executor_queue_length

Add new parameters for new features:

• --polling-threads (Can be used to specify specific threads only working on GPU polling)
• --print_times_per_timestep (prints additional timestep analysis after octotiger is done)

[diehlpk]

LGTM!