ENH: factor analysis Heywood cases

[josef-pkt]

I just ran the white wine example in Stata
https://gist.github.com/josef-pkt/1da753c160a054c9f342e735ee9be05e
which shows a Heywood case with one communality > 1

Stata reports, when using ipf

Beware: solution is a Heywood case
(i.e., invalid or boundary values of uniqueness)

Variable         |  Factor1   Factor2   Factor3 |   Uniqueness 
         density |   1.1050   -0.1253   -0.4318 |     -0.4231 

values are close enough, but not the same between Stata and statsmodels, at around 2nd decimal

Stata defaults to maxiter=1, which has high agreement between Stata and statsmodels, identical at print precision in summaries (4 decimals) for unrotated loadings.
At maxiter=1, there is no communality > 1

For MLE, Stata warns again about Heywood case, but sets uniqueness to zero

(tests formally not valid because a Heywood case was encountered)

Variable      |  Factor1   Factor2   Factor3 |   Uniqueness 
fixedacidity |   0.9994   -0.0353    0.0036 |      0.0000 
density       |   0.2992    0.9330   -0.2001 |      0.0000