Safe Reinforcement Learning via Curriculum Induction