A Representation Learning Perspective on the Importance of Train-Validation Splitting in Meta-Learning