Add linear mixed model to the differential expression interface

[Sizerta]

PR Checklist

• Referenced issue is linked
• If you've fixed a bug or added code that should be tested, add tests!
• Documentation in docs is updated

Description of changes

Closes #614
This adds pt.tl.LinearMixedModel to the DE interface.
It wraps statsmodels' MixedLM and fits one mixed model per variable with a random intercept. This means you can use all your cells while accounting for the donor they came from, or use random effects in pseudobulk-style designs.
The random-intercept syntax is the same (1 | variable) syntax already used by Milo, thanks to reusing parse_random_effects:

model = pt.tl.LinearMixedModel(adata, design="~condition + (1 | donor)")
model.fit()
res = model.test_contrasts(
    model.contrast(column="condition", baseline="A", group_to_compare="B")
)

# or let paired_by add the random intercept for you
pt.tl.LinearMixedModel.compare_groups(
    adata,
    column="condition",
    baseline="A",
    groups_to_compare="B",
    paired_by="donor",
)

Technical details

• A bit more honest about p-values: statsmodels uses a z-test for fixed effects, which is too optimistic when there are few donors. We use a between-within degrees-of-freedom correction instead. In null simulations with the condition assigned per donor (3 vs 3 donors), the z-test flagged 12% of genes at α = 0.05, the corrected test 5–6%, and Statsmodels without the random effect 57%. A calibration test is included.
• Less shouting about convergence: statsmodels can report non-convergence even when the optimizer has already reached the optimum (refitting changed the estimates by less than 1e-13). Those fits are retried with derivative-free optimizers unless method is explicitly provided. Any genuine remaining failures are reported once via a warning and the converged column.
• Small refactor: LinearModelBase.compare_groups now builds its formula through _comparison_design, which lets LinearMixedModel turn paired_by into a random intercept without changing the existing behavior of other methods.
• Adds 15 tests; the DE test suite passes (268 passed, 2 skipped).
• The scope is intentionally small for now: one random intercept. Random slopes, crossed random effects, and the MAST and DREAM options from the issue can come later.
• No new dependencies.
• Fitting takes roughly 20–30 ms per variable on 180 cells (single core) and supports n_jobs parallelization just like Statsmodels.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 97.27273% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 75.96%. Comparing base (54b9990) to head (393767b).

Files with missing lines	Patch %	Lines
...ifferential_gene_expression/_linear_mixed_model.py	97.05%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1147      +/-   ##
==========================================
+ Coverage   75.69%   75.96%   +0.26%     
==========================================
  Files          55       56       +1     
  Lines        8431     8537     +106     
==========================================
+ Hits         6382     6485     +103     
- Misses       2049     2052       +3     

Files with missing lines	Coverage Δ
src/pertpy/tools/__init__.py	91.42% <100.00%> (ø)
...py/tools/_differential_gene_expression/__init__.py	96.15% <ø> (ø)
...ertpy/tools/_differential_gene_expression/_base.py	96.06% <100.00%> (+0.12%)	⬆️
...ifferential_gene_expression/_linear_mixed_model.py	97.05% <97.05%> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.