Text-Guided Generation and Editing of Compositional 3D Avatars