West-of-N: Synthetic Preference Generation for Improved Reward Modeling