Add packaged JudgeArena task support for LUMI

[kargibora]

Summary

This PR adds support for packaged JudgeArena tasks in oellm-eval. JudgeArena remains responsible for task definitions, prompts, judging, and scoring; oellm-eval handles scheduling and result collection.

• Register JudgeArena tasks and language variants in the task groups.
• Prepare task inputs with judgearena tasks download.
• Run JudgeArena in the selected container or venv, with judge settings supplied through --judgearena_kwargs.
• Report failed runs and collect numeric metrics from JudgeArena’s run metadata.

Verification

I have verified this by reproducing 100-example COPA and BoolQ runs and 50-example Italian XCOPA which scored the same/after.

LUMI container

The LUMI base has been updated from ROCm 6.4 to ROCm 7.2.1 because the JudgeArena vLLM wheel requires ROCm 7.2 libraries. JudgeArena and vLLM should be installed in a separate virtual environment because they require a different version of torch. In the future, JudgeArena can also be added to other cluster containers.

---

Example

Run a packaged JudgeArena task with a candidate model and a local VLLM judge (given the models are already downloaded)

   oellm-eval schedule \
     --models Qwen/Qwen2.5-0.5B-Instruct \
     --tasks alpaca-eval-ja \
     --n_shot 0 \
     --limit 10 \
     --judgearena_kwargs '{"judge.model":"VLLM/Qwen/Qwen3-4B"}'

[kargibora]

I will mark this as a draft for more extensive testing, but it would be nice to hear some feedback because changing container version is not super straightforward and I have to fully verify nothing breaks..

[kargibora]

This is somewhat required to seperate environments. This doesnt change any old way of running the code.

[kargibora]

Currently the main does not have the required code for JudgeArena so I think as the first draft it is better to let this be some reference to a specific working commit