Learning Local Advantage Functions for Generalizable Graph Optimizations