Provide arithmetic NaN handling

[ErikMcClure]

inNative provides correct canonical NaN propagation, but currently does not attempt to produce arithmetic NaNs from floating point operations. Instead, it simply produces whatever NaN the platform feels like. At some point, this should at least be made an option under a flag, but it's not clear how many programs actually care about the type of NaN they get. Because nobody pays attention to the bit pattern of NaNs in most programs, this is considered a low priority fix.