Cut the Deadwood Out: Post-Training Model Purification with Selective Module Substitution