All Roads Lead to Likelihood: The Value of Reinforcement Learning in Fine-Tuning