Optimizing Language Models for Inference Time Objectives using Reinforcement Learning