Store every target as a list (target_is_scalar flag) so all splits load; to_task() in k_prompts.py

K Program Environment: train (2,900), eval (220), eval_depth4 (100), interpreter, prompts