REF: cleanup wald test contrast df_xxx

Bad naming !

ContrastResults uses df_denom for the p-value
self.pvalue = self.dist.sf(self.statistic, df_denom)

This should be a serious bug. The relevant lines show that they have test coverage.
(This got me worried for some time. test failure in statsmodels.base.tests.test_generic_methods.TestGeneric....test_ftest_pvalues if I make a change to df in chi2, no failure in regression or discrete tests, i.e. only conistency checks with verified pvalues)

However, df_denom in chi2 is not the same as df_denom in the F-test. It's the same as df_num.

https://github.com/statsmodels/statsmodels/commit/dbbf05a02d9490fe4c33d3ef0bf5c4fef9350db0#diff-b165b4bd4e10edd0dfb485e47c562b2cL1518

my problem: I never know which is the numerator and which the denominator
switch to generic df_constraints instead of df_num

aside: df_constraints for wald test are defined in terms of shape. Using rank should be safer. Also we don't check for "estimability" which I still haven't fully figured out.

Start with ContrastResults and the linked commit, then inspect the Wald-test degrees-of-freedom handling in the chi2 path. Run statsmodels.base.tests.test_generic_methods.TestGeneric....test_ftest_pvalues and the regression or discrete tests mentioned in the issue. Done means the degrees-of-freedom names consistently describe their use and the existing verified p-value checks still pass.