When resample (and benchmark?) get instantiated resampling instance, they should check validity

[mb706]

The following should not work, since the resampling instance references rows that are not contained in the task any more:

rs <- rsmp("holdout")
ts <- tsk("iris")
rs$instantiate(ts)
ts$filter(c(1, 51, 101))
ts
#> <TaskClassif:iris> (3 x 5): Iris Flowers
#> * Target: Species
#> * Properties: multiclass
#> * Features (4):
#>   - dbl (4): Petal.Length, Petal.Width, Sepal.Length, Sepal.Width
resample(ts, lrn("classif.rpart"), rs)
#> INFO  [16:34:57.794] [mlr3] Applying learner 'classif.rpart' on task 'iris' (iter 1/1)
#> <ResampleResult> with 1 resampling iterations
#>  task_id    learner_id resampling_id iteration     prediction_test warnings
#>     iris classif.rpart       holdout         1 <PredictionClassif>        0
#>  errors
#>       0

Or do we actually ever need to refer to rows that were $filter()-ed out before? The code anticipates this problem in some places, but the following error can only be triggered if the entire data was actually replaced using cbind:

rs <- rsmp("holdout")
ts <- tsk("iris")
rs$instantiate(ts)
ts$filter(c(1, 51, 101))
ts$cbind(iris[1:3, 1:4])
resample(ts, lrn("classif.rpart"), rs)
#> INFO  [17:05:26.512] [mlr3] Applying learner 'classif.rpart' on task 'iris' (iter 1/1)
#> Error: DataBackend did not return the queried rows correctly: 50 requested, 1 received.
#>         The resampling was probably instantiated on a different task.