adjust rounding decimals to expected values

This may have been the cause for some tinderboxes choking on this test
on Windows and Mac if the forecast results slightly differed due to the
initialization with random values and the calculation taking place with
a different internal floating point precision. So let's try.

Blindly throwing ROUND(...,12) at all values is wrong anyway because
that doesn't take the magnitude of the expected value into account. What
we actually want is an accuracy of 12 digits, not an accuracy of 12
decimals.

0.123456789012
1.23456789012
12.3456789012
123.456789012
...

Change-Id: If88d75277c3d17cfd72f36d0f88d093b63da8add