Investigate whether dispersion should be fixed or estimated from the data

[jr-leary7]

• related to Speed up GEE mode #241, as not estimating scale probably increases speed by reducing the number of scoring iterations
• in extreme cases, very high dispersion values lead to wildly inflated standard errors and thus deflated test statistics / weird plots
• dispersion is currently estimated via:

$$ \hat{\phi} = \left(-p + \sum_{i=1}^n n_i\right)^{-1} \sum_{i=1}^n \sum_{t=1}^{n_i} \hat{r}_{it}^2 $$

where $\hat{r}_{it}$ is the estimated residual for subject $i$ at timepoint $t$

[jr-leary7]

commit 990bee8 removed the gee.scale.fix argument in favor of strictly fixing dispersion to be equal to 1 throughout. this decision was based on benchmarking performed on simulated data, in which we saw that fixing the scale lead to slightly better dynamic gene classification as well as lower runtimes