mPOLICE: Provable Enforcement of Multi-Region Affine Constraints in Deep Neural Networks