Better behaviour when expected disagreement is 0 for Krippendorff's alpha

[logological]

Some measures of interannotator agreement, such as Cohen's kappa and Krippendorff's alpha, are not meaningful for cases where all annotators assigned the same label. In this case the expected disagreement is 0 and, were the measure's usual formula to be calculated, it would produce a division by 0.

When KrippendorffAlphaAgreement is told to calculate Krippendorff's alpha for such cases, calculateAgreement() returns 0.0. I think this is misleading—a value of 0.0 means that the agreement was successfully calculated, and that it was indistinguishable to a random labelling. I think that calculateAgreement() should have instead signalled to the user that it could not return a meaningful result with the data it was given. CohenKappaAgreement correctly handles this case by throwing an exception of type InsufficientDataException—maybe KrippendorffAlphaAgreement should do likewise…?

[logological]

Hmm… I see now that CohenKappaAgreement is throwing the exception via CodingAgreementMeasure. So the problematic behaviour described above is probably a consequence of KrippendorffAlphaAgreement not implementing ICodingAgreementMeasure, as mentioned in Issue #17.

[logological]

Looking at other IAA tools such as ReCal seems to confirm that Krippendorff's alpha is undefined when the expected disagreement is 0. I will push a change that checks getCategoryCount() and throws an exception if it's not 2 or more.