Create a forked KV IO transformer for exporting coreML Llama (pytorch#7778)

Summary:

Create a forked KV IO transformer for exporting coreML Llama
- As discussed in the group chat "
Design: KV cache IO on ANE: llama_transformer vs. static_llama", we agreed that forking a KV IO version llama_transformer is best for code quality and coreML/ANE development purposes

Differential Revision: D68424012

Author: Di Xu (SWE)