Add MMLU benchmark evaluation to evals

[CarlG0123]

Summary

• Add mmlu_eval() function for evaluating models on the MMLU (Massive Multitask Language Understanding) benchmark
• Add make_mmlu_data_loader() for loading MMLU data from HuggingFace (cais/mmlu dataset)
• Add MMLU_SUBJECTS (all 57 subjects) and MMLU_ANSWER_LETTERS module-level constants
• Uses standard MMLU zero-shot letter-prediction format: show all choices (A-D) in prompt, compare log probabilities for each answer letter token

Changes

• transformer_lens/evals.py: Added MMLU_SUBJECTS, MMLU_ANSWER_LETTERS, make_mmlu_data_loader(), and mmlu_eval()
• tests/acceptance/test_evals.py: Added 5 tests covering data loading (single/multiple/invalid subjects) and evaluation

Testing

• Verified with multiple models: Qwen1.5-7B (~40%), Mistral-7B (~45%), GPT-2, Pythia-1.4b
• Results align with published zero-shot MMLU benchmarks for these model sizes
• All existing tests continue to pass

Test plan

• pytest tests/acceptance/test_evals.py -v
• Verify MMLU data loading for single and multiple subjects
• Verify evaluation produces reasonable accuracy scores

🤖 Generated with Claude Code