Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment