AD index calculation fails for Mordred features

[paulsonak]

Mordred feature calculation can result in NaN's. DeepChem appears to use weights to zero these out, but we can't calculate AD_index without fixing the NaN's.

[paulsonak]

All-nan columns are excluded, which can result in problematic behavior. For example, in the original training dataset, a set of columns is all-nan while in a prediction dataset a different set of columns is all-nan. Then, the AD index can't be calculated because the shapes of the data are different for new vs original data.

I propose to no longer exclude all-nan columns for AD index calculation but instead change all-nan columns to all-zero columns, and continue imputing the mean for individual nan's in otherwise OK columns

@stewarthe6 @mcloughlin2 do you have a better general solution? should we dig further into what is causing all-nan columns during mordred feature calculations?