Add unweighted chi squared to CurveFitResult

[wshanks]

What is the expected behavior?

For #782, I was thinking about using something like the wonkiness criteria described in #417 but that requires the chi squared without weighting the data points by their error. Could we add the unweighted chi squared to the fields of CurveFitResult? What I want is typical deviation of a data point from the fitted curve.

@nkanazawa1989

[nkanazawa1989]

I dug into LMFIT implementation. They internally create Minimizer object to run fit, and this class implements own residual computation logic (chisq is computed from the output of this method). This logic internally delegates residual computation to an LMFIT model, and this method call can take user provided Minimizer options. Note that the Model._residual method that computes residual take weights argument and weights from data uncertainty are not applied unless this is provided. So the important point is whether the curve analysis calls the minimizer with weights argument, and at present this value is not supplied.
https://github.com/Qiskit/qiskit-experiments/blob/1137029588a4f3fc4a414e16077da7cb25ab426a/qiskit_experiments/curve_analysis/curve_analysis.py#L443-L450
This means CurveFitResult.chisq is unbiased. The data uncertainty is only considered in the cost function.
https://github.com/Qiskit/qiskit-experiments/blob/1137029588a4f3fc4a414e16077da7cb25ab426a/qiskit_experiments/curve_analysis/curve_analysis.py#L418-L423
Note that you can still bias chisq value by providing weights in your fit options, i.e.

    def _generate_fit_guesses(
        self,
        user_opt: FitOptions,
        curve_data: CurveData,  # pylint: disable=unused-argument
    ) -> Union[FitOptions, List[FitOptions]]:
        user_opt.add_extra_options(weights=curve_data.y_err)

        return user_opt