Learning Disentangled Avatars with Hybrid 3D Representations