Minor updates from tests on large fitting problems. Most `initialize` methods now account for masked pixels making them more robust. `Group_Model` objects will update the targets of their constituents if their target is updated. Edge case for LM is handled where it exits cleanly if asked to optimize a model with no free parameters.
The most visible update is that when plotting a `model_image` the models are not evaluated on the whole image, but instead only in their fitting windows. This is a better fit for the "principle of least surprise" in python. Users will expect these plots to be consistent with other parts of the code which only use the fitting windows.