From Scaling to Structured Expressivity: Rethinking Transformers for CTR Prediction