ENH: add cov_type fixed_scale to linear model

[josef-pkt]

closes #2142 (I opened that issue for the record and to link to the mailing list discussions)

cov_type that allows fixing the scale in the linear model.
main use_case is when weights in WLS are actual inverse variances

see discussions on mailing list

Todo:

• check other results
  • llf uses MLE scale
  • AIC, ... ?
• docstring
• example
• unit tests for scale kwd

DONE except for llf/loglike which needs refactoring, in all 3 linear regression models

Questions:
Should we change the default to use_t=False? answer on mailing list: Yes
I think so, but IIRC we don't do it now as part of cov_type. (use_t was part of cov_type but is now mostly a separate, independent keyword.)
correction: we already have use_t=False as default for every cov_type except 'nonrobust', set in RegresssionResults.__init__ - nothing to do in this PR

Do we want to overwrite the scale attribute?
It's not necessary for cov_params and a bit of a pain to change - issues for unrelated refactoring.
skip this for now. this depends on other uses of the scale attribute

Follow-up question

Where do we put the relevant chisquare gof test, which only applies in the case of fixed scale?

[coveralls]

Coverage increased (+0.01%) when pulling bca64ba on josef-pkt:cov_type_fixedscale into 923aaf3 on statsmodels:master.

[cdeil]

Thanks for putting this in!!!

As a unit test ... how about taking the example from scipy.optimize.curve_fit with absolute_sigma=True?
https://github.com/scipy/scipy/blob/7750f4f6e9f4a6fd7b026f658aab6cb72da2c249/scipy/optimize/tests/test_minpack.py#L341

[coveralls]

Coverage increased (+0.02%) when pulling 309f447 on josef-pkt:cov_type_fixedscale into 923aaf3 on statsmodels:master.

[josef-pkt]

@cdeil I added one set of unit tests from the scipy curve_fit tests. Thanks for the suggestion.

[josef-pkt]

I would like to leave the loglike refactoring to another PR, which means loglike and llf are not correct for fixed scale.

merge as is ?

[cdeil]

Looks good to me: 👍 to merge.

[josef-pkt]

only two merges behind master, merging as is

[josef-pkt]

@ceil Thanks for the feedback
I will come back to this, but for now I'm doing a round of unrelated merges.

that llf/loglike are independent of the cov_type is "not nice"