Towards Scalable Physically Consistent Neural Networks: an Application to Data-driven Multi-zone Thermal Building Models