Closing the Approximation Gap of Partial AUC Optimization: A Tale of Two Formulations