Revisiting Implicit Differentiation for Learning Problems in Optimal Control