Enhancing Size Generalization in Graph Neural Networks through Disentangled Representation Learning