POLICE: Provably Optimal Linear Constraint Enforcement for Deep Neural Networks