Expected data were re-written by using LibreOffice 5 as a reference.
We kept only 8 decimals, because this is the precision used in tests,
and it's easier to read without scientific notation.
Also a few expected values were wrong so they were changed. The correct
expected values were confirmed by both LibreOffice 5 and wolframalpha.com.
Finally fixed an actual bug in code were a NaN value were returned instead of
the correct `#NUM!`.
3a38cc8f0a