Bump llama-cpp-rs to 0.1.126

[AsbjornOlling]

...does what it says on the tin.

[AsbjornOlling]

A bunch of breakage here... Android, MacOS and Windows builds all fail.

[AsbjornOlling]

Android:

I originally introduced aarch64-linux-android builds on utilityai/llama-cpp-rs#605
This PR includes a cfg macro thingy that uses a different type on android targets than others.

In a later PR, utilityai/llama-cpp-rs#629 , this manual CChar definition was replaced by the more proper std::ffi::c_char

The recent PR utilityai/llama-cpp-rs#861 introduces a cstr_to_string function, that works on *const i8 instead of the platform-appropriate std::ffi::c_char. I'll try and make a PR to llama-cpp-rs that updates this.

[AsbjornOlling]

The android fix PR is up: utilityai/llama-cpp-rs#872

Pointed this repo towards the branch on our fork of llama-cpp-rs, just to show that it actually does fix the android builds.

[AsbjornOlling]

MacOS:

It seems like that macos build failure likely wasn't caused by changes in llama-cpp-rs, but rather by some funky stuff with our github actions caching (which wasn't doing much anyway).

@s194042 ran into the same issue and removed the macos caching in commit 3c1d8a5

Rebasing this branch on main should do it.

(regardless, the caching-induced failure seems flaky- can't easily reproduce)

[AsbjornOlling]

Windows:

The windows thing was the latest (and maybe last? 🎉) chapter of a long-standing issue.

It's a mess of nested CMakeLists.txt files, that include each other via ExternalProject_Add.
Having not completely understood the issue, it turns out that setting CMAKE_GENERATOR to Ninja solves the problem. I'm happy with that. See #255 for a short history.