enable parallel prefill (#4068)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/4068

CHanges include
- Model export changes to allow indexing into freqcis
- xnnpack delegate change to partition with dynamic shape
- runner change to do perfill in single step
ghstack-source-id: 231807015

//unrelated failures
bypass-github-export-checks

exported-using-ghexport

Reviewed By: larryliu0820

Differential Revision: D58874164

fbshipit-source-id: 61005137b90744d5a5b1a76d081f66c05d4dda8e