Update on "[ExecuTorch][Llama] Split custom sdpa op and kv cache"

Summary:
This enables us to do more easier module swap with model definitions
from torchtune

Test Plan:
CI

Reviewers:

Subscribers:

Tasks:

Tags:

[ghstack-poisoned]