This version splits the expectation data into sgl, dbl, ext format tests. The source operands and results for each use corresponding data sizes.
The tests are not completely symmetric - the sgl and ext tests implement fewer cases than dbl. There are a few reasons for this (some cases are less meaningful to duplicate - but time spent and space consumed are also factors...)
The fbcc tests are omitted from ext currently... If this is really important I can add it - but it's more fiddly to do than the others so I skipped it for now. Only 'brand' test variants are executed for ext format - mainly because I haven't generated tables for the others. 'brand' tends to find more errors anyway.
All trig ops have been placed under a conditional build flag, and are currently off, to save space.