How Important is the Train-Validation Split in Meta-Learning?