ENH: determine which initialization schemes cause `partial_fit()` and `fit()` to differ

[eiviani-lanl]

It'd be good to investigate the specific cases in which partial_fit diverges from full fit in coefficients. So far, I've encountered this primarily with the 'range' weight scheme and highly redundant input data. Given that the model appears fairly stable across a range of parameter configurations, I suspect the differences are more strongly influenced by input data characteristics than by the model parameters themselves.