LHM-Humanoid: Learning a Unified Policy for Long-Horizon Humanoid Whole-Body Loco-Manipulation in Diverse Messy Environments