Replace eval() in LooGLE evaluation parsing

[fallintoplace]

Summary

• replace eval() in the LooGLE cloze metrics with safe JSON/literal parsing
• reuse the same safe parser for qa_pairs when building the LooGLE Hugging Face dataset
• add regression tests for valid Python-style literals and non-literal payloads

Root cause

The LooGLE benchmark parsed both dataset fields and generated answers with eval(). For shortdep_cloze, that meant a crafted prediction string could execute Python during metric calculation.

Validation

• uv run --python 3.11 pytest tests/test_loogle_parsing.py
• uv run --python 3.11 flake8 evaluation/benchmarks/loogle/calculate_metrics.py evaluation/benchmarks/loogle/create_huggingface_dataset.py evaluation/benchmarks/loogle/parsing.py tests/test_loogle_parsing.py
• uv run --python 3.11 mypy evaluation/benchmarks/loogle/calculate_metrics.py evaluation/benchmarks/loogle/create_huggingface_dataset.py evaluation/benchmarks/loogle/parsing.py tests/test_loogle_parsing.py --check-untyped-defs
• uv run --python 3.11 python -c "import evaluation.benchmarks.loogle.create_huggingface_dataset as mod; print("import-ok", hasattr(mod, "build_task_dataframe"), hasattr(mod, "main"))"

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.