Curvature-Regularized Variational Autoencoder for 3D Scene Reconstruction from Sparse Depth