implement penalized likelihood ratio tests to calculate p-values

[jzluo]

In logistf, likelihood ratio test statistics are not calculated by excluding the parameter being tested when fitting the null model. Instead the coefficient is constrained to zero such that it still contributes to the penalization (see https://stats.stackexchange.com/questions/555971/calculation-of-p-values-in-logistf-package)

In the logistf code, it seems this is done in the calculation of XW^(1/2), going by where selcol appears. So I implemented a masking in _get_XW().

There's some minor discrepancies in the p-values compared to logistf. Need to look into that but it's fine for now.

package	Intercept	age	oc	vic	vicl	vis	dia
logistf	8.020268e-01	6.143472e-03	8.751911e-01	1.678877e-06	1.237805e-05	5.449701e-02	4.951873e-03
firthlogist	7.66583794e-01	6.11138932e-03	8.26365363e-01	1.67218633e-06	1.23618050e-05	5.34899080e-02	4.84687363e-03