Add mask filter option to MapFit

[adonath]

This PR implements the handling of an exclusion mask in the MapFit object. I tried a few possible implementations, such as applying the exclusion mask to the stat array and in the total_stat() method.
In the end the most efficient solution was to apply the mask directly to the exposure map and setting the exposure of the corresponding pixels to zero. The downside is, that the exposure map is now different for MapFit and MapEvaluater. There the question whether MapEvaluater should take an exclasion mask as well. I was sure whether this makes sense, but maybe it does?

[cdeil]

This is in response to #1633 , so @registerrier - you might also want to chime in.

@adonath - This is only for a spatial mask at the moment? Or does it work for a 3D mask or to select energy range as well?

Is applying the mask only to exposure OK? I guess cash is the only stat we use at the moment, and indeed I see that it is cash=0 if mu_on=0. But it might happen due e.g. to FFT noise from the PSF convolution that this results in slightly non-zero mu_on and weird stats effects? And what about background: it's a bug to just set exposure and npred from source counts to zero, but to keep positive background and n_on, no?

Maybe we could look to Sherpa for guidance and see what they do with masks passed via "filter" or "notice" calls?

[cdeil]

The downside is, that the exposure map is now different for MapFit and MapEvaluater. There the question whether MapEvaluater should take an exclasion mask as well. I was sure whether this makes sense, but maybe it does?

People are using the MapEvalutor to get model and residual maps. So I think we have to add the same mask option there as well.

[adonath]

@cdeil The mask is ND now, so it can be used to filter the energy axis or other axes as well.

In general I agree this requires some more discussion, as it was not clear to me how to implement this best (and correct) either. So should the mask only be applied to the stat array? Or even only to the total_stat sum? Should maps such as exposure and background and npred be masked as well?
Yes, maybe it's a good idea to check how sherpa implements this.