Differentiate between initial split and splits in resamples in the accessor functions `training()`/ `testing()` / `analysis()`/`assessment()`

[hfrick]

TMwR calls the elements of the initial split "training" and "testing" set and the elements of the a resample "analysis" and "assessment" set.

In rsample, both of those types of splits are rsplit objects and all of training()/ testing() / analysis()/assessment() work on all rsplit objects. We should be more consistent with that conceptual difference and make training()/ testing() work only on initial splits and analysis()/assessment() only on splits in resamples.

library(rsample)

split_init <- initial_split(mtcars)
cars_train <- training(split_init)
cars_analysis <- analysis(split_init) # this should error

folds <- vfold_cv(mtcars, v = 3)
fold_1 <- folds$splits[[1]]
analysis_1 <- analysis(fold_1)
train_1 <- training(fold_1) # this should error

^{Created on 2022-12-06 with reprex v2.0.2}

[hfrick]

A follow-up question to ponder, possibly its own issue: Do we want any of these functions to work on rset objects?

This is motivated by some of those sets only containing one split anyway (the ones for a validation split/set and the apparent set). For example, for a validation set, it feels sorta natural to want to be able to access the one analysis (aka training) data from the rset.

For those sets containing multiple splits, analysis() and assessment() could gain an index argument similar to get_rsplit(x, index). training() could return the $data of the splits.