FLamE: Few-shot Learning from Natural Language Explanations