Feature/multi dataset support

[Henry-Shan]

Integrate ARC, HellaSwag, WinoGrande, TriviaQA, and HLE datasets

Add support for 5 new benchmark datasets into the RouterArena evaluation pipeline.

Key changes:

Add evaluation configs (config/eval_config/zero-shot/):

• ARC-Challenge.json, ARC-Easy.json (mcq_accuracy)
• HellaSwag.json (mcq_accuracy)
• WinoGrande.json (mcq_accuracy)
• TriviaQA.json (triviaqa_exact_match)
• HLE.json (exact_match)

Create data preparation scripts (scripts/):

• prepare_arc_data.py: Load ARC-Challenge/Easy from HuggingFace,
  normalize numeric answer keys (1/2/3/4 → A/B/C/D), combine all splits
• prepare_hellaswag_data.py: Load HellaSwag, convert label indices to
  letters, skip unlabeled test split
• prepare_winogrande_data.py: Load all 6 WinoGrande subsets, handle
  binary choice format
• prepare_triviaqa_data.py: Load all 8 TriviaQA subsets, save full
  alias lists for flexible answer matching
• prepare_hle_data.py: Load HLE test split, filter image-based questions,
  handle both exactMatch and multipleChoice answer types

Update evaluation pipeline (llm_evaluation/run.py, evaluate_models.py):

• Add split handling in load_ground_truth_dataset() for all 5 datasets
• Add dataset name detection in evaluate_models.py
• Add triviaqa_exact_match metric (alias list matching) to metrics.py
• Register triviaqa_exact_match in eval_reasoning.py

Update inference pipeline (router_inference/generate_prediction_file.py,
llm_inference/run.py):

• Add new splits to DATASET_PATHS and filename routing
• Add split support for router predictions and argparse choices

[yl231]

Apologies for the long silence, @Henry-Shan — maintainer triage. This is a sizable change (~992 lines across 15 files) for multi-dataset support and predates several months of pipeline evolution, so it no longer applies cleanly to main. Before investing in a full review: is this still active, and is it aimed at the current RouterArena evaluation flow? If yes, a rebase onto the latest main would let us review it properly; if it's been superseded, let us know and we'll close it. Thanks for the contribution either way.