Beyond Independent Genes: Learning Module-Inductive Representations for Gene Perturbation Prediction