Add option to bias-correct GEE variance-covariance matrix for small samples

[jr-leary7]

• as per Li and Redden (2014), using the standard sandwich variance-covariance matrix to perform a Wald test of significance - as we do in waldtestGEE() - leads to inflated Type I error (false positive) rates when the number of subjects is small and / or the number of timepoints per-subject is large
• this was seen in our simulation study, where we observed high numbers of false positives when using $n_s = 6$ subjects
• scRNA-seq datasets quite often have low numbers of subjects compared to observations per-subject, making this a problem for the GEE mode
• several small-sample bias corrections exist, the easiest of which to implement is the DF correction, where $p$ is the number of covariates:

$$ \mathbf{V}_{df} = \left(\frac{n_s}{n_s-p}\right) \mathbf{V} $$

• thus, it makes sense to add an option to bias-correct the estimated sandwich variance-covariance matrix when the number of subjects is lower than some heuristic threshold e.g., $n_s = 30$

[jr-leary7]

• in addition to the naive DF correction mentioned above, implement the Kauermann and Carroll correction which should work better (i hope)
• this is much more complicated, as it involves recreating the subject-level correlation ($\mathbf{R}_i$) and vcov ($\mathbf{V}_i$) matrices as well as the projection matrix $\mathbf{H}$
• probably need to built out bias correction into its own function
  • call it biasCorrectGEE() and do not export to users
  • arguments: fitted model object, data to fit model, assumed correlation structure (AR-1 or exchangeable), and variable specifying subject ID