Pre-train, Align, and Disentangle: Empowering Sequential Recommendation with Large Language Models