Trading Carbon for Physics: On the Resource Efficiency of Machine Learning for Spatio-Temporal Forecasting