Is Meta-training Really Necessary for Molecular Few-Shot Learning ?