[etLLM] KV Cache as IO by default, as attribute optional

🚀 The feature, motivation and pitch

For accelerators (HTP and ANE), KV Cache as IO is used to better fit accelerator needs and performance optimization. However, in llama_transformer, KV cache is embedded as module attribute, which is also the running mode for CPU. It brings deviation in the entire flow. For example, we may maintain unnecessarily different runtime logics for different backends.
This issue is to set KV Cache as IO by default, as attribute optional.

• It does not affect CPU performance.
• Exposing KV cache as IO would help user manage KV cache.
• Keep the attribute option for specific use cases where KV cache may not be passed by IO, for example, in Vulkan.

Alternatives

No response

Additional context

No response

RFC (Optional)

No response

cc @mergennachin @cccclai @helunwencser @jackzhxng