Causality Meets Locality: Provably Generalizable and Scalable Policy Learning for Networked Systems