resampling

[m-muecke]

resampling methods to add:

• time series holdout
• time series rolling window cross-validation
• time series expanding window cross-validation

further todos:

• documentation for holdout
• documentation for cross-validation

Notes:

• perhaps split cross-validation into expanding/rolling window and sliding/fixed window cross validation since both are in forecast_cv
• the current implementation is based on sorting the ids of the task, will have to see if this logic holds otherwise I see two options: either the user is on his own and data is already sorted or better sort the data based on the date column, perhaps as well on a key index in case that will be supported. Perhaps only support this in the as_task_fcst() functions and no the more low level once.

PR of old mlr3 implementation that is based on an ordered task with date column: mlr-org/mlr3#633