Towards the Pedagogical Steering of Large Language Models for Tutoring: A Case Study with Modeling Productive Failure