pnm: Optimize and improve validation for read_separated_ascii

[mstoeckl]

Per commit message:

This function now rejects numbers that start with a leading +, although to match netpbm behavior it does not reject leading zeros. It also avoids allocating on the non-error path, reducing the time needed to read a large plain PPM test image to 50%. As the function no longer keeps the entire token in memory at once, the error messages no longer specify the bad token.

1. I don't think losing error message specificity here is a big problem, because PNM images are essentially never written by hand so the errors will rarely occur; if parsing does fail, one can easily find non-digit non-whitespace characters in a file with a regexp (e.g. grep -P '[^0-9 \t]')
2. To create a large "plain" (P3) image to test this with, use e.g. pngtopnm -plain large.png > ~/output.ppm. Depending on the png bit depth this can create both 8-bit and 16-bit PPM files.

[fintelia]

Thanks!