Comparing Bad Apples to Good Oranges: Aligning Large Language Models via Joint Preference Optimization.