Validation split as a 3-way split

[hfrick]

This PR is the start of introduction functionality to work with a split into training/validation/testing as a 3-way split as opposed to two consecutive binary splits (testing vs not-testing and then not-testing into training and validation). Closes #369.

As an illustration of where this is going:

library(rsample)
data(ames, package = "modeldata")

# in this PR
first_split <- initial_validation_split(ames)

# in the next PR
rset_for_tuning <- validation_set(first_split)

# access to individual subsets 

# in this PR
ames_train <- training(first_split)
ames_val <- validation(first_split)
ames_test <- testing(first_split) 

# in the next PR
ames_analysis <- analysis(rset_for_tuning$splits[[1]])
ames_assesment <- assessment(rset_for_tuning$splits[[1]])

I'll comment on the code for specific things I'd like feedback on.

In addition to that: Do we need methods for obj_sum() and type_sum() for initial_validation_split objects? I assume they are only needed for usage with pillar / within tibbles and we don't do that for initial splits?

Edit: Don't merge this yet, I'm collecting changes in this branch.

[hfrick]

this is analogous to the single-element vector specifying the proportion for training in initial_split(), here taking two of the three proportions (training and validation).

[topepo]

Maybe this should wait until we have a function to extract a validation split set object from this...

Can you add a details section describing that this is going to be the preferred methods for this splittings strategy and show the two-step code using initial_split()/validation_split() sequence compared to what we suggest doing now?

[hfrick]

yes, that sounds good! Once we have the function to make the rset, I'll update the docs here and for validation_split() showing the code for a 3-way split vs two consecutive binary splits.

[github-actions]

This pull request has been automatically locked. If you believe you have found a related problem, please file a new issue (with a reprex: https://reprex.tidyverse.org) and link to this issue.