On the Complexity of First-Order Methods in Stochastic Bilevel Optimization