Harnessing deep learning for physics-informed prediction of composite strength with microstructural uncertainties