compile time string value of NaN should be float.nan instead of nanF

[dkorpel]

Came up in #23268

dmd's floatToBuffer formats nan values as follows:

pragma(msg, float.nan.stringof); // nanF
pragma(msg, double.nan.stringof); // nan
pragma(msg, real.nan.stringof); // nanL

This leads to confusing error messages and invalid header generation:

float n = nanF;

It should instead use something that's valid D syntax.

[crazymonkyyy]

pragma(msg,long.init.stringof);