Beyond Uniform Scaling: Exploring Depth Heterogeneity in Neural Architectures