Fix floating-to-integral narrow undefined behavior

[Carson Radtke (carsonRadtke)]

Summary

• reject undefined floating-to-integral conversions before narrow_cast
• preserve exact signed minimum conversions across differing floating-point precision
• cover open range boundaries, infinities, NaN, and bool
• add Linux ASan/UBSan and Windows ASan CMake presets and CI jobs

Testing

• WSL Clang C++20 ASan/UBSan: 15/15 tests passed
• Windows MSVC C++20 ASan: 15/15 tests passed
• Windows MSVC C++20 Debug utils_tests: passed

Fixes #786