Interpretable Lightweight Transformer via Unrolling of Learned Graph Smoothness Priors