DSpAST: Disentangled Representations for Spatial Audio Reasoning with Large Language Models