[RF] Weighted Datasets for RooKeysPdf including negative weights

[MindaugasSarpis]

Up until now, weights were ignored when computing kernel width in RooKeysPdf.

This was worked on for 6.28, but negative weights are causing evaluation errors (some weights can go negative, like sWeights).

It would be good if negative weights were allowed and cases, where they are not allowed, would throw exceptions rather than computation be attempted.

Relevant forum post:
https://root-forum.cern.ch/t/roofit-keyspdf-fails-in-6-28-works-in-6-26/54440