Physics-informed neural networks for non-Newtonian fluid thermo-mechanical problems: an application to rubber calendering process