Fix use of MPI GPU RDMA variables

[msimberg]

The variables were inconsistently named, leading to the CMake variable not working. I've also made the CMake variable dependent on DLAF_WITH_HIP. At this point I've simply made them consistent, but I'd also like to propose a rename of the variable since it applies to both HIP and CUDA.

Some options:

• DLAF_WITH_MPI_GPU_SUPPORT
• DLAF_WITH_MPI_GPU
• DLAF_WITH_GPUDIRECT
• DLAF_WITH_GPU_RDMA
• etc.

I think I quite like the first option as it's quite straightforward to understand and uses generic enough terms (e.g. I just learned that GPUDirect seems to be an NVIDIA marketing term: https://developer.nvidia.com/gpudirect).

[RMeli]

+1 for DLAF_WITH_MPI_GPU_SUPPORT

[msimberg]

cscs-ci run

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests :white_check_mark:

Comparison is base (ea4521d) 94.05% compared to head (9a872ee) 94.04%. Report is 10 commits behind head on master.

:exclamation: Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

```diff @@ Coverage Diff @@ ## master #1088 +/- ## ========================================== - Coverage 94.05% 94.04% -0.02% ========================================== Files 148 148 Lines 9204 9204 Branches 1165 1165 ========================================== - Hits 8657 8656 -1 Misses 325 325 - Partials 222 223 +1 ```

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.