$\Text{d}_{2}\text{o}$: Dynamic Discriminative Operations for Efficient Long-Context Inference of Large Language Models