Updates for llama.cpp and ggml

[Christian Kastner <ckk@debian.org>]

Hi,

I just uploaded ggml 0.9.4, and llama.cpp b6643 will follow tomorrow
morning [1].

The dependency on ggml is still pretty tight. For example, llama.cpp
b6653 from only a day later did not build with ggml 0.9.4. I can be
assumed that ggml will be bumped eventually, but just to show that this
is still under quite active development.

The new llama.cpp also improves upon the autopkgtests. The tests need
non-free models [2] to run, and while downloading them is an option,
it's not a performant one. Instead, I've provided a way for testbeds to
pass in this data. The tests will run if they find models they expect,
and will otherwise be skipped. I'll share more details and examples
later on.

Best,
Christian

[1]: Otherwise llama.cpp's Salsa jobs get upset because the required
     ggml isn't available yet.

[2]: At least, that is the current assumption, even if the model
     weights are released with a DFSG-compatible license. This is
     a very restrictive interpretation.