BUG: disallow misleading cov_type HCx in models other than linear regression models

In GLM and discrete Models, all hcx return the same cov_type, which (IIRC) is HC0 (of HC1 ?), i.e. standard Huber-White sandwich without any corrections for hatmatrix, ...

Based on a twitter thread, I found this blog https://datacolada.org/99 Blog post points out that HC3 is much better than HC1 in OLS.

We might be able to add the hatmatrix correction to GLM and so switch to proper HCx. (I don't remember having seen that in the literature for GLM.)

First step add warning if HCx, x>0 is used. (Need to verify which one. I usually use HC0 for plain sandwich).

(Related, I recently looked briefly at cluster robust CRx again, similar to bias-reduced CR in GEE)