Machine learning for alkali-activated concrete: feature attribution, strength–carbon relationships, and the limits of out-of-campaign generalisation