When Can Proxies Improve the Sample Complexity of Preference Learning?