Take dynamic shape as an argument to LLMEdgeManager and (#4241)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/4241

remove unnecessary dynamic shape

As titled.

If we don't have `use_kv_cache` enabled, or we have
`enable_dynamic_shape`, the input token dimension is dynamic. With
`enable_dynamic_shape` enabled, we don't need to set dynamic shape for
input_pos since it will always be a scalar tensor.

Reviewed By: tarun292

Differential Revision: D59759976

fbshipit-source-id: 8740a196fe6929f8839e7a6d107f1cf51211ab7d