This also makes us a bit more accurate, due to better rounding of
intermediate results.
This also gives us the flush-to-zero and denormals-are-zero SSE settings
for free! (Assuming UE is build with SSE)
Long doubles are always at least 80 bits wide in memory and it suffices
if we can address these 80 bits, to mark the long double as NAN at the
end of an MMX instruction, so the additional magic using conditional
types is unnecessary.
