Improving 3D-aware Image Synthesis with A Geometry-aware Discriminator