This commit tries to fix the discrepancies between the results of
floating point calculations during runtime and compile time, see
e.g. fptools/ghc/tests/numeric/should_run/arith008.hs. Part of the
story was the fact that floating point values are represented as
Rationals in GHC and therefore never lost precision, at least for the
operations for which constant folding is done. To compensate for this,
before and after floating point operations the operands are
temporarily converted to/from Float/Double. This is wrong, because
host architecture and target architecture are confused this way, but
in a non-cross-compiling scenario it works.