DARIL: When Imitation Learning outperforms Reinforcement Learning in Surgical Action Planning