Data-Efficient Learning via Clustering-Based Sensitivity Sampling: Foundation Models and Beyond