Background computation in cataloging is not sufficiently robust?

[schlafly]

When running the mosaic level pipeline on the files in
/grp/roman/eschlafly/misc/background-consistency/
via something like strun roman_mos blah.json and via strun romancal.step.SourceCatalog i2d.asdf I get substantially different number of sources. It looks like a few things contribute here:

• the image that the mosaic pipeline sees is not identical to the image the source sees, because the source catalog step adds and subtracts a background. This makes almost no difference but does impart a small amount of numerical noise.
• somehow the RomanBackground is rather sensitive to this numerical noise and finds a significantly different answer between the two images
  RomanBackground should be robust to this---it looks to be intending to use a median over a large area. We need to figure out what's going on here so that it is actually more robust.

We should also think about:

• there are a handful of bad discrepant points, probably due to poor linearity correction / inverse linear correction in the reference files. These should be fixed but also shouldn't cause issues with the background.
• We are adding & subtracting to the model.data. That saves some memory but seems like bad form. We should probably just making a copy of this plane (roughly 64 MB for a skycell), so that the cataloging step does not change the input image at all.