Null-Space Constrained Low-Rank Adaptation for Response-Specified Large Language Model Unlearning