ENH: Adding a sampling strategy to set the total number of samples to be used for all classes, using class proportions

TEST: Deactivate baseline comparison for Shark RF method, because shark uses a random generator for which the seed can not be set. Hence, the model will be slightly different at each run

ENH: Ensure RAM parameter is used by persistent filters (should be applied likewise for all applications)

The "count-insert" idiom is an antipattern that leads to searching the
element twice. Avoid that and also the 0 comparison a bit below.

In addition, cache the end iterators of the two maps. This is safe as
the iterators stay valid when the map is modified.