tolerate differences in test ouput using numdiff

Using numdiff command (available on naf-xfitter), it tolerates differences smaller than 1e-4 (either relative or absolute) between any numbers having the same position in two files. One can think of tuning this tolerance for each test, if needed.

It works for me for all tests except profilerLHAPDF (there different numbers of eigenvectors are produced).

Do you think it is acceptable solution for us? Do you have numdiff on your machine? (if numdiff is not available, the usual diff is used and some tests will fail)