Add Mind2Web evaluation task with two candidate pool sizes

Add Mind2Web web navigation task for ACE framework:
- mind2web: ~200 candidate elements per step (199 negative + positives)
- mind2web2: ~50 candidate elements per step (49 negative + positives)

Each task includes:
- prepare_data.py: Downloads Mind2Web from HuggingFace, converts to
  step-level ACE samples with candidate element selection formulation,
  performs stratified train/val/test split by domain
- data_processor.py: Three-level evaluation (element index + operation
  type + value matching) with flexible parsing
- run.py: Standard ACE training/evaluation script with offline, online,
  and eval_only modes
- data/sample_config.json: Data path configuration

The two versions enable studying the effect of candidate pool size on
ACE's context learning performance for web agent tasks.