Split degrees of freedom and variance estimation

[graemeblair]

this would allow lm_robust to match difference_in_means

[lukesonnet]

I think that splitting out variance from degrees of freedom is a good goal. For example, the only reason we aren't applying the Bell-McCaffrey DoF correction to HC2 is because that isn't standard in other packages.

Whereas with CR2, no one else implements it besides clubSandwich, so we apply to correction so that we match clubSandwich.

I don't know if we do this for this next launch and I'm tagging this as estimatr 0.8.0

[lukesonnet]

TODO: write down the full matrix of estimators we want to support