Add more tests to sdpa_with_kv_cache for speculative decode (#4325)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/4325

Simulate speculative decode with next iter larger than one token decode
bypass-github-export-checks

ghstack-source-id: 234820963

Reviewed By: manuelcandales

Differential Revision: D60045219

fbshipit-source-id: 8034fd234c7ccf648d27a3a8b57025c751ade3eb