Update benchmark scripts: basic separation for alps vcluster, number of cores, gpu defaults, ...

[albestro]

Main discussion points:

1. create a "base" alps-gh200 config so that each vcluster can inherit and customise it
2. reduce number of cores: see dedicated comment below
3. is pika mpi mode 31+1 the best one? should we go for this or keep the default initialised (30+0)? @msimberg @biddisco
4. currently we cannot set differently DLAF params for MC and GPU, and for GPU defaults are not the best. should we set it like this or do we prefer to have them in the generator scripts as cli parameters?

• TFactor: bulkerify GEMV (both MC and GPU) #1214 required for DLAF_TFACTOR_NWORKERS in point 4

[albestro]

About number of cores to use, there was the question if using all cores might have been problematic, e.g not leaving enough resources to the system.

In some cases it revealed to be problematic (e.g. due to an OpenMP problem that seems to use 1 thread more than requested), so here it is a quick benchmark.

Next are time plots (the lower the better) for eigensolver. The naming of series are distinguished by two numbers:

• first number is how many cores the allocation is given (-c slurm parameter)
• second number is how many pika threads are used (--pika:threads parameter)

These benchmarks ran all on current master (0.7.0) with --dlaf:red2band-panel-nworkers=50.

From a quick look I wouldn't say there is any real difference, except on --local runs which seem to have a small benefit having more cores available (72-72 > 72-64 > 64-64).