cohen's kappa can't handle single label cases

cohen_kappa_score assumes that the data is going to have multiple labels, although most time's that is the case for training a classifier that may not always be the case for testing, like for example if you are just evaluating a classifier over many datasets and one of them only has one class data. Sure this evaluation is kind of useless but this will happen when running a classifier through unknown datasets where you are trying to estimate classification performance.

Here how the problem could be reproduced:

from metrics import cohen_kappa_score
cohen_kappa_score([0,0],[0,0])

The above results in a nan, when in reality it should throw an error, or return a 1 with a warning. The one is because after all the classifier we are evaluating in this case predicted perfectly.

The entry point is cohen_kappa_score; first run the supplied two-line reproduction and inspect its implementation and related metric tests. Resolve the expected single-label behavior—an error or a warning with a result of 1—and add regression coverage for the chosen outcome.