Beyond Deep Learning Dominance: A Scaffold-Aware Hybrid Framework for Robust Toxicity Prediction in Data-Scarce Regimes