From Large to Small Datasets: Size Generalization for Clustering Algorithm Selection