This change makes the localgeodesicfefunction test fail again.
So there must be a bug in the corresponding non-finite-difference
code.  Looking at simplex cases confirms me in this opinion.
But I don't find the bug...

[[Imported from SVN: r8463]]