Not All Preference Pairs Are Created Equal: A Recipe for Annotation-Efficient Iterative Preference Learning