Make std::chrono::duration formatting more similar to C++20 standard

[matrackif]

Hello,

Currently std::chrono::duration with the %S and %T specifiers is not printed according to the spec. This issue was addressed but later dropped in #2207. Many commits have appeared since then so I have tried to implement a new approach.

Writes the second as a decimal number. If the number of seconds is less than 10, the result is prefixed with 0.
If the precision of the input cannot be exactly represented with seconds, then the format is a decimal floating-point number with a fixed format and a precision matching that of the precision of the input (or to a microseconds precision if the conversion to floating-point decimal seconds cannot be made within 18 fractional digits). The character for the decimal point is localized according to the locale.

Now, the %S specifier only prints with millisecond precision. I have implemented the subsecond extractor, inspired by the MSVC STL implementation of std::format for the hh_mm_ss class. Please see the new unit tests.

Further comments:

1. The decimal point is not formatted according to the locale
2. The %T specifier now behaves according to the C++20 standard and should be equivalent to %H:%M:%S
3. Possible but highly unlikely overflow in the subsecond_helper functions. Not sure if fmt_safe_duration_cast is required and more analysis might be needed.
4. Floating point subseconds are truncated by casting to std::chrono::duration::period to std::uintmax_t, this behavior is probably acceptable according to the spec as it only requires the value to be as precise as the the std::ratio in std::chrono::duration::period . The MSVC implementation uses std::floor

Comments and suggestions are welcome.

Thanks

[matrackif]

@vitaut Hi do you mind running the workflows one more time for this PR? Thanks

[vitaut]

Thanks for the PR but I don't think it can be merged in because it incorporates code from other projects and would require license updates.

[matrackif]

@vitaut Thanks for your reply,

I believe Howard Hinnant's MIT license is compatible with the current license. Otherwise, if you would like to see this feature make it in then let me know. We could devise a similar but more original solution in a different PR and resolve issue #2207

Regards

[vitaut]

The problem is not license compatibility (although {fmt}'s license is substantially different in that it only requires source attribution) but the fact that these are separate licenses so we would need to include both. The two options are:

1. Writing the suggested improvement from scratch.
2. Asking the original code copyright owner (Howard?) if they are OK with distributing a portion of their code under the {fmt}'s license.

Since the changes are relatively simple I suggest going with option 1.