util: Adjust Fraction class addition overload overflow tests

[jamescowens]

@Tahvok, the arch maintainer, caught a subtle error in the extreme overflow tests for the newly augmented Fraction class in the 5.4.6.0 release. Arch, unlike other distributions, uses the C++ compiler flag -D_GLIBCXX_ASSERTIONS, and this caught the error.

Essentially the problem is in the fraction addition overflow tests at the extreme negative case. When the fraction addition overload operator goes to add fractions of the same denominator, where both addends are std::numeric_limits<int64_t>::min() / 2, the result on the numerator will be std::numeric_limits<int64_t>::min(). When simplification is called, it will call std::gcd, which calls std::abs. The std::abs of std::numeric_limits<int64_t>::min() is an overflow, because it would be std::numeric_limits<int64_t>::max() + 1, but this is handled as an assert in glibc. So to prevent this, we tightened the negative side overflow constraint in overflow_add by 1 to ensure the numerator will never be less than std::numeric_limits<int64_t>::min() +1, so that the std::abs of that will succeed (i.e. will be std::numeric_limits<int64_t>::max()).

This is really esoteric and will not be reached by any actual wallet operation, but it is being treated as a hotfix to get Arch compiles to work.

Thanks to @div72 for helping me troubleshoot this.

[div72]

tACK 1269669.

[jamescowens]

This PR has been merged also into master and tagged as 5.4.6.0-hotfix-1.