DeCo: Decoupling Token Compression from Semantic Abstraction in Multimodal Large Language Models