[TBD] Review batching for GPU jobs

[tcompa]

For GPU jobs, we should turn off batching (as in "multiply CPU&memory requirements by the number of concurrent tasks") and only support internal queues (as in "I need to pack 5 identical tasks in a same SLURM job, and they will run sequentially")

Note: in principle we could also drop the srun xx & srun yyy & wait pattern, and move to simply listing the commands to be run. But not having sruns any more would be a bit annoying because we use them when we parse&analyze Fractal SLURM usage.
TBD.