Cleanup NAN code in float.h, too.

 Cleanup NAN code in float.h, too.
 
 In the same spirit as 3bf905692, assume that all compilers we still
 support provide the NAN macro, and get rid of workarounds for that.
 
 The C standard allows implementations to omit NAN if the underlying
 float arithmetic lacks quiet (non-signaling) NaNs.  However, we've
 required that feature for years: the workarounds only supported
 lack of the macro, not lack of the functionality.  I put in a
 compile-time #error if there's no macro, just for clarity.
 
 Also fix up the copies of these functions in ecpglib, and leave
 a breadcrumb for the next hacker who touches them.
 
 History of the hacks being removed here can be found in commits
 1bc2d544b, 4d17a2146, cec8394b5.
 
 Author: Tom Lane <tgl@sss.pgh.pa.us>
 Discussion: https://postgr.es/m/1952095.1759764279@sss.pgh.pa.us