The Goldilocks Paradigm: Comparing Classical Machine Learning, Large Language Models, and Few-Shot Learning for Drug Discovery Applications