Training and Testing with Multiple Splits: A Central Limit Theorem for Split-Sample Estimators