KR test seems sensitive to permutations of rows of data

[mcdonohue]

See tests in gls branch where KRmodcomp results were different for different permutations of the data. Tolerances were increased to allow tests to pass, but results should probably be invariant to these permutations.