Move CI to RHEL9 and CUDA 12.1

This MR moves the Allen CI to RHEL 9 and CUDA 12.1.

• Uses the CERN provided docker image for alma linux 9 gitlab-registry.cern.ch/linuxsupport/alma9-base in build jobs
• Uses the target x86_64_v3-el9-gcc12+cuda12_1 from LCG103.
• A small reference update due to the change of compiler to CUDA 12.1 was necessary.

Will come at a later stage:

• There are no clang12 platforms in LCG 103 yet, so for the moment we will stick to gcc12 and gcc12+cuda12_1 builds.
• A new HIP installation is in the works, we will use HIP 5.4.3 once it is available in CVMFS.

The impact in speed is the following:

Device                        Throughput (kHz)    Reference Throughput (kHz)  Speedup    % change    Status
--------------------------  ------------------  ----------------------------  ---------  ----------  ---------
NVIDIA RTX A5000                        120.37                        123.18  0.98x      -2.28%      OK
AMD EPYC 7502 32-Core                    15.31                         16.56  0.92x      -7.54%      DECREASED
NVIDIA GeForce RTX 2080 Ti               97.73                         99.85  0.98x      -2.12%      OK
NVIDIA GeForce RTX 3090                 156.05                        154.88  1.01x      0.76%       OK

It is clear that the impact in CPU performance comes from moving from clang to gcc. In the GPU case this would have to be studied better.

The simplified Allen CI passes. The full Allen CI passes with one test exception, see issue #413 (closed)