fix parseNumber returning zero for finite values with negative exponent

[sage-mode-hunter]

Reading a numeric string as a number, for example doc["x"].as<double>() on
the value "1" followed by 300 zeros and e-30 (which is 1e270), returns 0.

1. the exponent overflow guard adds the parsed exponent to exponent_offset even when the exponent is negative, so a large positive offset combined with a negative exponent trips the guard and yields 0 instead of scaling the mantissa back down.
2. saturating at exponent_max is also too tight once the mantissa carries extra digits, which is why the check now uses the reach of the binary powers-of-ten table.

Compute the effective exponent with its real sign and only saturate to inf/0 past what the table can represent; multiplyByPowerOfTen() handles everything in range, including subnormals.

[bblanchon]

Hi @sage-mode-hunter,

Thank you for this contribution.

This is an interesting fix, but is it really useful?
How did you end up with a one followed by 300 zeros?
I understand that, in theory, this is a valid input, but so is a one followed by one billion zeros, so why bother?

I looked at your user profile, and I found it rather peculiar.
The account was created only 3 weeks ago and is opening numerous pull requests to unrelated projects.
It claims to work at a company that has absolutely zero online presence, except for an AI-generated website registered 3 weeks ago—what a coincidence!

All of this is very suspicious and makes me question the real intent of this PR.

Best regards,
Benoit