AnchorDS: Anchoring Dynamic Sources for Semantically Consistent Text-to-3D Generation