GaussianFormer: Scene as Gaussians for Vision-Based 3D Semantic Occupancy Prediction