L1: Controlling How Long A Reasoning Model Thinks With Reinforcement Learning