MaskedUprootTree only applies its mask to `pandas.df` not `array` or `arrays`

This isn't an issue in any existing analysis that I am aware of, but it is an unintended inconsistency and hence I consider it a bug. The only stage that I know of in current use is the fast_carpenter.BinnedDataFrame stage which uses pandas.df to get data, and therefore histograms appear masked.

This could be a useful feature in some sense, since it allows expressions and user stages to succeed event selections without issue, since they actually see all events. However the inconsistency in the interface is risky, so I think this should be made more explicit.