Neural Light Field Estimation for Street Scenes with Differentiable Virtual Object Insertion