A machine learning perspective on the inverse problem indentation: uniqueness, surrogate modeling, and learning elasto-plastic properties from pile-up