Separate the Wheat from the Chaff: Model Deficiency Unlearning via Parameter-Efficient Module Operation