[R-package] sorting of observation weights and initial scores is broken in lgb.cv()

[jameslamb]

In #2503 , it was reported that using certain sampling strategies could cause incorrect results if the indices for subsetting weren't sorted. This was "fixed" in the R package by #2524 , however that PR introduced a bug noted in #2524 (comment).

Essentially, slice.lgb.Dataset() takes in a vector of indices (row numbers) and returns a new Dataset with that subset. On its way to doing that, it sorts those indices. If a model had observation weights or initial scores, they don't get sorted the same way, meaning they'll refer to the wrong observations.

To fix this, need to ensure that tuples of indices, weights, and initial scores are all kept together when re-sorting indices.

[rgranvil]

@jameslamb Is this still being worked on? Anything I can do to help? I'm anxious to get this fixed :)

[jameslamb]

@rgranvil yes it's still being worked on, sorry for the delay! I hope to have this fixed w ithin the next day

[jameslamb]

@jameslamb Is this still being worked on? Anything I can do to help? I'm anxious to get this fixed :)

Hey @rgranvil , I think I have this working! It will be a bit of time until reviewers are able to give feedback on my latest revisions and merge them, but if you're ok with living on an unmerged version of the R package then you can try the fix I just pushed to #2573 .