feat: use apptainer backend in grpo training

[andre15silva]

On Alvis, it's a good idea to set APPTAINER_CACHEDIR and APPTAINER_TMPDIR to not run out of space. In Berzelius, I think only the APPTAINER_CACHEDIR was needed, but I need to check once we get the project again.

[andre15silva]

[andreafo@alvis3-41 CodeRepairRL]$ uv run python test_nano_batch_timing.py 
👋 This is mini-swe-agent version 1.17.0.
Loading global config from '/cephyr/users/andreafo/Alvis/.config/mini-swe-agent/.env'
Testing batch size 2
  Run 1: 56.28s
Average time for batch size 2: 56.28s

@BjarniHaukur is this an acceptable duration for running training? This script is running the same thing as in the nano_agent.py file (under __main__)

[andre15silva]

Pushed the files I'm using on Alvis so that you can replicate on your side