splits based on twinning

[topepo]

Twinning is a tool to create splits based on making the marginal distributions of the variables the as close as possible. See this paper (pdf) for more details.

There is an R package that can be used.

We could use this in vfold_cv() as well as initial_split() and mc_cv().

[hfrick]

Adding some notes from a slack discussion:

• If this also applies to the group_*() versions but not the time_*() version, it would be good to have that as an argument rather than new functions.
• We could use this as a multi-variable strata solution; if someone passes a single column to strata, the splits are made as they are now. If 2+ columns are chosen, we could use twinning to do the stratified splits.