Add support for sampling with the combination of splitratio and replace = FALSE

[theo-s]

No description provided.

[edwardwliu]

Related to sampling, it looks like when applying the doubleBootstrap for honest random forests without groups and folds, we ignore observationWeights and uniformly sample with replacement (see code) for the averaging set. When groups and folds are used, it looks like we sample with replacement according to observationWeights (see code). As a corollary, we also always take a doubleBootstrap when using groups/folds.

1. Is this understanding correct?
2. If so, is this intended behavior?

[theo-s]

I believe this understanding is correct.

Right now this is intended behavior, but I could see reason to change this to use the weights in the second bootstrap as well. The only complication would be that the meaning of the weights would change based on the sample that is taken in the first bootstrap (since those observations are removed from sampling and their weights are removed from the total set of weights that would be used in the next step).