formatted_size() is broken for modern compile-time API

[alexezeder]

There is a problem using formatted_size() with format strings for compile-time API - FMT_COMPILE and _cf. Here Is an example - https://godbolt.org/z/zz3c1E, and a simplified version of it - https://godbolt.org/z/drW1rG.

This problem can be solved in at least two ways:

1. By adding SFINAE check to

   fmt/include/fmt/compile.h

   Lines 837 to 840 in ab0f7d7

   	template <typename CompiledFormat, typename... Args>
   	size_t formatted_size(const CompiledFormat& cf, const Args&... args) {
   	return format_to(detail::counting_iterator(), cf, args...).count();
   	}

   so it would not be available for compiled_string-based format strings. Thus the version from core.h will be chosen.
2. By adding a new formatted_size() implementation specifically for compiled_string-based format strings. This also requires marking a conversion operator in compiled_string-based format strings explicit, which results in changing few lines in compile.h for conversion (with using "custom" FMT_STRING_IMPL macro).

There is no difference between these 2 ways in performance because of compilers optimizations. But there is a difference between runtime API and compile-time API (lack of support for named spec'ed fields, for example). So which way should we choose to fix broken formatted_size()?

Maybe it's not even a problem because, maybe, a format string from compile-time API shouldn't be used in formatted_size().

[vitaut]

I think we should go with the second option except that instead of making the conversion operator explicit the formatted_size overload should be templatized like other formatting functions.