Check if `shap` accepts `pd.DataFrame` (and for which Explainers)

[hbaniecki]

...and possibly remove the conversion from pandas to numpy in ShapWrapper.

Then, explainer.model_info['arrays_accepted'] would become redundant, which in fact fixes https://github.com/ModelOriented/DALEX/issues/507 by making the scikit Pipeline fail gracefully in the downstream package (with an appropriate exception message).