converted fractional_part_rounding_thresholds to use an array of u32s

[Cazadorro]

converted fractional_part_rounding_thresholds to use an array of u32s instead of a unicode literal character array for compatibility with CUDA and non CUDA platforms, based on discussions in #4167

[Cazadorro]

Note that this works with vcpkg as well (tested with overlay port), ran tests locally all appeared to pass (note however I'm on MSVC 2026). I did use the special CUDA compiler arguments for UTF8 in the separate project I used to test VCPKG compatibliity, dealing with that is a whole other can of worms I didn't want to include in this fix, I just wanted to make sure I could use the escape hatch with those compiler arguments as with out this change you can't use a fmt at all no matter what you do with vcpkg and cuda.

below are the arguments I used in the separate project

target_compile_options(cuda_experiments PRIVATE
        $<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=-utf-8 -Xcudafe=--unicode_source_kind=UTF-8 -Xcicc=--unicode_source_kind=UTF-8>)