Analyzing Koopman approaches to physics-informed machine learning for long-term sea-surface temperature forecasting