DARTS: Decoder-Aware Representation Tuning via Surgery for Model Merging