Better handling of NaN (not a number)

[Even Rouault]

• It appears that with Microsoft C-Runtime, atof("nan") returns 0 whereas it returns a nan number in Linux. printf("%f", a_nan_floating_point) can return "1.#QNAN", "-1.#QNAN" or "-1#.IND" on Windows, whereas it always returns "nan" on Linux. So the solution is to update CPLStrtodDelim() to handle the "nan" string specially, and be carefull when turning a value into a string by testing with CPLIsNan(). Several places such as gdal_translate, gdalwarp utility, geotiff, PAM and VRT nodata reading&writing needs to be updated in consequence.

• We also need special code to handle the case where nan is the nodata value when comparing a pixel value to the nodata value, such as GDALWarpNoDataMasker(), GDALNoDataMaskBand class, VRT complex source. This is also the opportunity to extend the use EQUAL_TO_NODATA() that was introduced in r19671

[Even Rouault]

Fixed in trunk (r19692). Tests added in r19693