of
        CSR samples, that is, a CSR sample should contain (i) rows
        that are all filled, (ii) rows with no elements at all, and
        (iii) rows that are partially filled. At the same time and for
        the given total number of specified elements (nnz), there
        should be minimal preference to rows with a given number of
        elements.  To achieve this, the sampling method is built-up on
        using a sawteeth model for counts. In the simplest case, we
        would have

          counts = arange(n_rows + 1) % (n_cols + 1)

        that has equal number of all possible column counts per row.
        This formula can be used only for specific input values of
        n_rows, n_cols, and nnz. To generalize this model to any
        combinations of inputs, the counts model above is extended
        with an incomplete sawtooth, and the right and lower
        rectangular parts that will guarantee that

          counts.sum() == nnz

        for any combination of n_rows, n_cols, and nnz. Basically,
        we'll find a maximal window in (n_rows + 1, n_cols + 1)-grid
        that is able to hold a sequence of sawteeth and so-called
        final correction, while the external part of the window is
        filled with counts to meet the nnz contraint exactly.
        r