Use the canonical #include location for json.hpp

[cebtenzzre]

The two simplest ways to get nlohmann/json are to install it system-wide, or to vendor it via CMake.

System

If you install it system-wide, it will be at /usr/include/nlohmann/json.hpp (on Arch Linux, at least). minja is not compatible with this out-of-the-box, requiring you to add the non-standard #include path of -I/usr/include/nlohmann.

This is confirmed by /usr/share/pkgconfig/nlohmann_json.pc, which does not list /usr/include/nlohmann as a required include directory:

Name: nlohmann_json
Description: JSON for Modern C++
Version: 3.11.3
Cflags: -I/usr/include

CMake

If you provide the json dependency as part of a CMake project like this:

add_subdirectory(deps/json)  # or FetchContent_Declare/FetchContent_MakeAvailable
target_link_libraries(mytarget PRIVATE nlohmann_json::nlohmann_json)

The same problem occurs, requring a manual target_include_directories(mytarget PRIVATE /path/to/nlohmann). See here.

---

This PR fixes those issues by using the canonical header location. This is consistent with the nlohmann json documentation.

[cebtenzzre]

cc @ochafik

[ochafik]

Thanks @cebtenzzre , this is also consistent with moving minja down to vendored folder in ggml-org/llama.cpp#11900, I guess the obvious follow up will be to move json.hpp down in llama.cpp soon too.