Add an option to enable --runtime-repack in llama.cpp

[ekcrisp]

Is your feature request related to a problem? Please describe.
After updating to 0.3.4 ARM optimized Q4_0_4_4 models are no longer supported by llama.cpp. Instead when loading an error is thrown that "TYPE_Q4_0_4_4 REMOVED, use Q4_0 with runtime repacking". This comes from this change in llama.cpp. I believe a flag must be added to llama-cpp-python and passed down llama.cpp internal to enable this feature.

Describe the solution you'd like
Add a flag for Llama instantiation that enables runtime repack in llama.cpp

[ekcrisp]

After looking at the llama.cpp PR it seems like you just need to set this cmake flag to on, GGML_CPU_AARCH64 but when I tried doing CMAKE_ARGS="-DGGML_CPU_AARCH64=ON" pip install llama-cpp-python --upgrade --force-reinstall --no-cache-dir the performance of the Q4_0 was far worse than Q4_0_4_4 gguf before the update, and nothing in the verbose model output suggests that this was applied.

[ekcrisp]

This may be a bug in llama.cpp. I will close this ticket if repack works after this issue is closed upstream