Towards scalable physically consistent neural networks: An application to data-driven multi-zone thermal building models