A Two-Stage Active Learning Algorithm for $k$-Nearest Neighbors