Statistics for 3D-SSGAN: Lifting 2D Semantics for 3D-Aware Compositional Portrait Synthesis