NaN #defines

[Zack Weinberg]

We currently define NAN to be a single precision quiet NaN.  All
other unqualified math #defines are doubles; this is an odd
inconsistency.  Also, there's no way to get a signalling NaN.  Perhaps we
should extend the definitions?

NAN	double, quiet
NANf	single, quiet
NANl	extended, quiet

NANs	double, signal
NANfs	single, signal
NANls	extended, signal

Also, bits/nan.h isn't being included by math.h -- is this intentional?

zw