[ot][spam][ai] POLICE: Provably Optimal Linear Constraint Enforcement for Deep Neural Networks