Add *some* hardness model for gradient minimization exercise

[amkrajewski]

One idea is to use one of @dovahkiin0022 models from out ULTERA papers. We should be able to just convert to ONNX and throw it in quickly.