correct coeftable for GLMMs, add tests

[dmbates]

See #307

Note that the example for this PR ("goldstein") still has the peculiar property that the deviances change between the Laplace approximation and nAGQ=11 but the parameter estimates are exactly the same. Now the multipliers, m11.mult are all close to 1.0 but not that close.

I don't think this is right still

[codecov]

Codecov Report

Merging #308 into master will increase coverage by 0.27%.
The diff coverage is 92.50%.

@@            Coverage Diff             @@
##           master     #308      +/-   ##
==========================================
+ Coverage   95.69%   95.97%   +0.27%     
==========================================
  Files          20       21       +1     
  Lines        1510     1515       +5     
==========================================
+ Hits         1445     1454       +9     
+ Misses         65       61       -4     

Impacted Files	Coverage Δ
src/MixedModels.jl	100.00% <ø> (ø)
src/generalizedlinearmixedmodel.jl	83.33% <83.33%> (+2.26%)	⬆️
src/linearmixedmodel.jl	100.00% <100.00%> (ø)
src/mixedmodel.jl	100.00% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 872bba1...753747e. Read the comment docs.

[dmbates]

It looks as if the identical parameter estimates for the Laplace approximation and the nAGQ=11 fit for the goldstein data may be a coincidence. Fitting such models to the contra data does produce slightly different values for the parameter estimates.

[palday]

We're now missing StatsBase.vcov for GeneralizedLinearMixedModel and I'm wondering if it makes sense to gather up all the methods defined for the superclass MixedModel into one place.

[dmbates]

@palday Will you have the opportunity to look at the latest commit to see if it addresses the questions you had?

[palday]

There are a couple of spots where the superclass methods aren't there yet (coeftable that depends on coef, which isn't defined for GLMM), but I'll add them in.

[palday]

Ah, can't do it for the cherry picking after all -- the fixef methods for GLMM and LMM take different arguments (GLMM has a permuted argument, but LMM doesn't -- it instead depends on the difference between fixef and coef). So that would be a breaking change. I'll spin that re-factoring off into a different PR.

[palday]

Ah, so good news -- the coeftable method for GLMM is actually correct in the last release. The wrong table was introduced in #287 when the static allocation for the coefficients and errors was introduced for LMM. It still makes sense to add in the GLMM parallels for those things, then we can go back to more generic methods for things like coeftable that build upon them.