hyperband and multiple splits / resampling

[berndbischl]

maybe we can simply, in each race, increase the number of holdout evaluations? this would work singlecrit and also multicrit.

multicrit this would be a bit like a "rolling tide" thing

if we do this, we have to calculate it in a way that each bracket still takes the same amount of time

yes this is nearly the same issue as #5 but as the discussion there went into a different direction i opened a new one

[jakob-r]

Please check discussions in #56 and #5 and continue here.