Enhancing and Accelerating Large Language Models via Instruction-Aware Contextual Compression