Adjoint Method versus Physics-Informed Neural Networks in PDE-Constrained Inverse Problems