Better handling of constant response

[palday]

For GLMM, this means allow constructing a model with a constant response.
This is useful for simulation.
For both GLMM and LMM, catch an initial PosDefException.
If the PosDefException is from a constant response, throw ArgumentError.

Technically this is a "feature" so I've done a minor version bump instead of patch.

[codecov]

Codecov Report

Merging #578 (30fbe94) into main (9f27f2b) will decrease coverage by 0.02%.
The diff coverage is 90.00%.

❗ Current head 30fbe94 differs from pull request most recent head 8eca919. Consider uploading reports for the commit 8eca919 to get more accurate results

@@            Coverage Diff             @@
##             main     #578      +/-   ##
==========================================
- Coverage   96.25%   96.22%   -0.03%     
==========================================
  Files          28       28              
  Lines        2509     2516       +7     
==========================================
+ Hits         2415     2421       +6     
- Misses         94       95       +1     

Impacted Files	Coverage Δ
src/linearmixedmodel.jl	97.92% <80.00%> (-0.18%)	⬇️
src/generalizedlinearmixedmodel.jl	89.93% <100.00%> (+0.09%)	⬆️

---

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 9f27f2b...8eca919. Read the comment docs.

[palday]

The drop in coverage is from the rethrow() lines when the PosDefException occurs without a constant response. I don't have a good pathological example for eliciting that handy.

[dmbates]

Looks good. Thanks for doing this.

Well, except for that test error thing.

[palday]

@dmbates Yeah, it looks like subtle linear algebra differences lead to GLMM failing at an unexpected point for GLMM, so I've just moved to always testing for constant response in fit!(::GeneralizedLinearMixedModel). The test with any should exit pretty quickly for most datasets where the response is not constant; the big exception would be a dataset that had been sorted by the response ....

[palday]

(The test should still be faster than trying to add error trapping into the objective evaluation.)