Unmasking the false reliability: supervised feature importance methods lack consistency and dose-response validation in desalination research