Support llama.cpp's no KV cache offload option

[eckartal]

• Users need the ability to run models without KV cache offloading in llama.cpp
• Currently, this option is not available, limiting flexibility for those optimizing memory usage.

Context

Success Criteria

• Add support for no KV cache offload in llama.cpp.