From Experimentation to Prediction: Interpretable Machine Learning for Efficient Design of Compressed Earth Blocks