convert raw data with shape (subject, rater) to (rater1, rater2)
brings data into correct format for cohens_kappa
Parameters:  data : array_like, 2Dim
bins : None, int or tuple of array_like


Returns:  arr : nd_array, (n_cat, n_cat)

Notes
no NaN handling, delete rows with missing values
This works also for more than two raters. In that case the dimension of the resulting contingency table is the same as the number of raters instead of 2dimensional.