The current implementation/documentation was made to avoid sNaN because of
potential safety issues implied by old/bad LLVM documentation. These issues
aren't real, so we can just make the implementation transmute (as permitted
by the existing documentation of this method).
Also the documentation didn't actually match the behaviour: it said we may
change sNaNs, but in fact we canonicalized *all* NaNs.
Also an example in the documentation was wrong: it said we *always* change
sNaNs, when the documentation was explicitly written to indicate it was
implementation-defined.
This makes to_bits and from_bits perfectly roundtrip cross-platform, except
for one caveat: although the 2008 edition of IEEE-754 specifies how to
interpet the signaling bit, earlier editions didn't. This lead to some platforms
picking the opposite interpretation, so all signaling NaNs on x86/ARM are quiet
on MIPS, and vice-versa.
NaN-boxing is a fairly important optimization, while we don't even guarantee
that float operations properly preserve signalingness. As such, this seems like
the more natural strategy to take (as opposed to trying to mangle the signaling
bit on a per-platform basis).
This implementation is also, of course, faster.
439576fd7b