restore and save optsum for GLMM

[palday]

closes #790

• add entry in NEWS.md
• after opening this PR, add a reference and run docs/NEWS-update.jl to update the cross-references.
• I've bumped the version appropriately

[codecov]

Codecov Report

Attention: Patch coverage is 97.43590% with 1 line in your changes missing coverage. Please review.

Project coverage is 96.93%. Comparing base (c0b9e2d) to head (61ed6f3).
Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
src/serialization.jl	97.05%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #791      +/-   ##
==========================================
- Coverage   96.94%   96.93%   -0.01%     
==========================================
  Files          33       33              
  Lines        3370     3397      +27     
==========================================
+ Hits         3267     3293      +26     
- Misses        103      104       +1     

Flag	Coverage Δ
current	96.64% <97.43%> (+0.05%)	⬆️
minimum	96.87% <97.43%> (-0.01%)	⬇️
nightly	96.57% <97.43%> (+0.05%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[ararslan]

I don't feel like I can give a super thoughtful review as I'm not familiar with this part of the codebase but here are some comments/suggestions in case they're useful, plus an approval since nothing seemed obviously wrong (as far as I could tell).

[ararslan]

I suppose it can never be the case that length(ops.lowerbd) < length(m.β) or that ops.lowerbd has indices that don't start at 1?

[palday]

lowerbd is in a struct we control and it's always Vector, so the indices are fine.

There are two cases here:

• the "slow" optimization, where the parameter vector contains both beta and theta, so length(lowerbd) needs to be length(beta) + length(theta).
• the "fast" optimization, where the parameter vector contains only theta, so length(lowerbd) needs to be length(theta). (In this case, there is an inner optimization step that finds the best beta for a given theta using PIRLS -- this will generally give the same answer or a very close one to doing free optimization over both, but there are a few pathological cases where the two methods can diverge, in which case the extra freedom of the slow optimization will yield more accurate results.)

The initial model creation initializes the lower bound to the length of theta because that is shared with the initialization for the linear mixed model (where only theta is included in optimization and beta is solved for directly).