Physics-informed self-supervised learning for predictive modeling of coronary artery digital twins