Fix missing llama2 model on Device Farm (#4162)

Summary:
This changes the input to upload the export llama2 model from S3, the caching logic will be implemented upstream in https://github.com/pytorch/test-infra/pull/5405

kirklandsign I see that running llama2 on S22 runs OOM now, do you have any insights on that?

Pull Request resolved: https://github.com/pytorch/executorch/pull/4162

Reviewed By: kirklandsign

Differential Revision: D59416499

Pulled By: huydhn

fbshipit-source-id: d5be31c7c0b14d6d16935ae0986195a39f7fe087