On the Effective Horizon of Inverse Reinforcement Learning