backend: rebase llama.cpp on upstream as of Sep 26th

[cebtenzzre]

Update our llama.cpp dependency, and update the backend code to be compatible with the changes:

• There are new translation units (cpp files) in llama.cpp, now added to our llama.cpp.cmake script.
• The model state serialization functions have been renamed, and they have a more robust API that does bounds checking and computes the state size exactly.
• The sampler API has been changed from a procedural API to a more object-oriented one, with a chain of sampler objects that is built in advance. (llama : refactor sampling v2 ggerganov/llama.cpp#9294)
• The log level constants have been renumbered for some reason. Now they are not in any particular order, and care must be taken in the log callback to check the log level of a message against each one individually, and handle GGML_LOG_LEVEL_CONT specially.

Support for the new model architectures is not included in this PR.

TODO: Update the changelog for the python bindings.