Behaviour of `ExcessMapEstimator`

Currently, the ExcessMapEstimator behaves differently for MapDataset and MapDatasetOnOff. In the first case, the npred is taken as the background - ie, the background and the source models are both considered. In the second case, only the counts_off and the background are considered, and the counts from the source ignored.

This is quite confusing. We should enforce similar behaviour, and I think that the npred should be considered in case of MapDatasetOnOff as well. What do you think @registerrier @adonath ?

[registerrier]

I think we should make things uniform with SpectrumDatasetOnOff where we have a npred_sig that focuses only on the model npred. And then indeed we could modify background so that it returns the background taking into account the model and possibly add an npred that would return background + npred_sig.

[adonath]

Here is the conclusion / action items how to unify the behaviour:

• Modify both SpectrumDatasetOnOff.backgroud and MapDatasetOnOff.background to return the expected (marginalised) background using get_wstat_mu_bkg()
• Introduce MapDatasetOnOff.npred_sig() and modify MapDatastOnOff.npred() to add the background component

I would proposed to add an optional mu_bkg argument to WStatCountsStatistic in a separate PR. I still need to make up my mind about the details…