T-TEST: Avoid NaN in paired-sample correlation significance calcuation.

Sometimes the correlation comes out just above 1.0 due to inaccuracy of
machine calculation, which causes sqrt(1-corr**2) to yield NaN.  This
commit forces the correlation into the valid range, fixing the problem.

Thanks to Matej Cepl <mcepl@redhat.com> for reporting the problem.
Thanks to John Darrington for suggesting the fix.