Use ast.literal_eval instead of eval in hellaswag_arabic_pfn

[jayzuccarelli]

Fixes #1293

hellaswag_arabic_pfn (the hellaswag_okapi_ar prompt function) parsed line["endings"] with the builtin eval(). That field is a string-encoded list coming from the OALL/AlGhafa-Arabic-LLM-Benchmark-Translated dataset on the Hub, so a crafted value would execute arbitrary code on the machine running the eval.

This switches it to ast.literal_eval(), which parses the same string-encoded lists and raises ValueError on anything that is not a plain literal. It's also the pattern already used elsewhere in the codebase for string-encoded dataset fields (e.g. sacrebleu.py). A repo-wide grep shows this was the only remaining eval() on dataset content.

Tests added in tests/unit/tasks/test_arabic_tasks.py:

• legitimate string-encoded endings still parse into the same Doc (choices/gold_index)
• a non-literal payload in endings raises ValueError instead of executing — this test fails on main (the payload runs) and passes with this change

pytest tests/unit/tasks/test_arabic_tasks.py: 2 passed. ruff check and ruff format --check clean on both touched files.