Deep Learning-based surrogate models for parametrized PDEs: handling geometric variability through graph neural networks.