Fixes gdiplus tests in Clang build.
When the compiler assumes non-trapping floating-point math, it may apply more aggressive optimizations. For example, in the fast path of `GdipInvertMatrix`, it may merge paired divisions into a single `divss` instruction. This can leave unused lanes whose results are ignored, but they still execute and may cause a division by zero, breaking tests that have the FP exception mask enabled.
Using `-ffp-exception-behavior=maytrap` prevents the compiler from performing such transformations.
It's a bit surprising that this flag also disables optimizations like those seen in the dmime tests, so there’s likely room for improvement on the compiler side. Regardless, relying on `round()` in tests isn't valid, as it's not available in msvcrt.dll and may depend on compiler-specific behavior.
-- v2: configure: Build PEs with -ffp-exception-behavior=maytrap. dmime/tests: Use ceil instead of round.