Two different methods for evaluating predictions

[MatteoRobbiati]

After a discussion with @BrunoLiegiBastonLiegi, we suggest to divide the one_prediction function into the following two:

1. prediction without mitigation inside;
2. mitigated_prediction with mitigation inside.

This can be useful for adding a new level into our analysis: also the parameter shift rule makes large use of the predictions (basically it evaluates the observable twice, using two shifted parameters $\mu^{\pm}$) instead of the original $\mu$. With the purposed splitting we can:

• perform the fit without mitigation;
• perform the fit with mitigation only at the end (executing predict_sample with mitigated_prediction);
• perform the fit with prediction while evaluating gradients and mitigated_predictions while calculating predictions;
• perform the fit using mitigated_prediction each time a prediction is called.

[MatteoRobbiati]

I think this can be closed, right?

[BrunoLiegiBastonLiegi]

Yes, even though, at the moment the default behaviour of vqregressor is the last one. The other possibilities are not implemented yet.