use C++23 #embed to speed up compile

[roj234]

Also easier to edit and keep git log small
Only tested on LLVM-MinGW
Require some refactor to action script

(cherry picked from commit be84a3d)

[roj234]

Current unable to compile on GitHub Actions...

[leejet]

This would introduce a dependency on Python. One of the reasons I converted merges/vocab into an .hpp file in advance was precisely to avoid bringing in extra dependencies.

[roj234]

Python is not necessary when compile using C++23, they are only used to re-create those hpps for compilers not support #embed yet (as as fallback), and GitHub Actions have Python installed.
If you don't like to remove those hpps, they can be included via macros.
On my computer, use #embed can save 15 seconds (about 1/3 time when I link to a pre-compiled GGML library). I feel the speedup and a cleaner git history are quite beneficial.
I'm not an expert on CMake and GitHub Actions, so I'm currently struggling with compiler arguments and CI failure.
To finish this PR, I may need to change CMake config and tune some compiler arguments... (However I have never used MSVC)

[roj234]

It can compile now, however no speedup at all, only a Python polyfill convert #embed to .h at runtime, some tweak on compiler options may be required.

[leejet]

Thanks for the PR. I’ll close this for now since it adds build/CI complexity without a clear measurable benefit. We can revisit once #embed is widely supported and dependency-free.