Leveraging sound speed dynamics in a generative deep learning framework for ray-based ocean acoustic tomography