Unify mask handling in Dataset methods and Estimators

We have to unify the handling of mask_safe and mask_fit in the dataset methods and Estimator classes. Here is the current behaviour:

Dataset methods:

• MapDataset.stack(), SpectrumDataset.stack(): takes into account mask_safe
• MapDataset.to_image(): takes into account mask_safe
• MapDataset.to_spectrum_dataset(): no mask handling
• MapDataset.resample_energy_axis(): takes into account mask_safe

Estimator classes:

• Fit: takes into account mask_safe as well as mask_fit
• FluxPointsEstimator: takes into account mask_safe as well as mask_fit
• LightCurveEstimator: takes into account mask_safe as well as mask_fit
• TSMapEstimator: no mask handling
• ExcessMapEstimator: takes mask_safe into account, optionally the mask_fit as well
• ExcessProfileEstimator: no mask handling (related to .to_spectrum_dataset())

Maker classes:

• FoVBackgroundMaker: just uses the exclusion_mask parameter, see #2921.

My current proposal goes into the direction of just applying mask_fit and mask_safe everywhere, except maybe for MapDataset.stack(), but I’ll think about once more. Please add your thoughts @registerrier , @AtreyeeS, @luca-giunti .

Related issues:

• #2953
• #2865
• #2701
• #2921

[adonath]

I’d be open to handle the mask_fit as well in dataset methods, it would unify the behaviour 100% and offers a bit more flexibility as @luca-giunti suggested…

[adonath]

From a quick check this looks all good now in Gammapy, except for one oddity: the ExcessMapEstimator exposes an option apply_mask_fit, which allows to control the behavior. None of the other estimators does this and one could easily achieve the same behavior using:

dataset.mask_fit = None

My recommendation would be to remove the apply_mask_fit option.