Fix weight handling in fitting

[nmearl]

Addresses #505.

[nmearl]

Reposting conversation from previously incorrect PR:

@nmearl:

@eteq Strangely, there's a difference when using the weights explicitly by passing in weights=1 / unc, and having them be implicit weights='unc' (where 'unc' means do 1 / spec.uncertainty.array). You can see the difference in value in the test...

@hcferguson:

@nmearl As I mentioned on some previous ticket....usually one wants the "weights" to end up being (uncertainty**-2). Perhaps it's doing that internally when you pass it weights='unc'?

It's sort of semantics...if chisq1 = sum(w * (predicted-model)**2) then you want w=uncertainty**-2. If chisq2 = sum(w**-2 * (predicted-model)**2), then you want w=1/unc. My impression is that chisq1 is more common practice when talking about the relation between weights and uncertainties.

@nmearl:

@hcferguson Ah, that's correct. The astropy fitters (e.g. the LevMarLSQFitter implements the chisq1 approach when using weights -- I'll fix the unc version accordingly.