When we use train_only_size > 0.0 in the leave_k_out_split function,
the list of unique users is thinned at the entrance to the _take_tails function.

Then in the _take_tails function, the bincount method produces zeros in the middle of the array and the `cumsum' method starts duplicating the ends.

For example (operation of the _take_tails function):
arr = [0 0 0 4 4 4]
np.bincount(sorted_arr) = [3 0 0 0 3]
end = [2 2 2 2 5]

As a result of the operation of the leave_k_out_split function, ratings multiplied by 4 will appear in the test for user 0

Must be:
arr = [0 0 0 2 2 2]
counts = [3 3]
end = [2 5]