Towards Physically Consistent Deep Learning For Climate Model Parameterizations