Task 2: one merged jsonl per split (captions + gold + image paths); drop baselines and per-file tsvs

Add Task 2 (CRAI-Bench): train/dev/test with captions, human gold labels, reference and generated images, organizer baselines