[Executorch][llama] Change runner to decouple prompt length from sequencelength

Following previous diff now we can utilize entire kv cache to generate more
tokens than max prompt length allowed.

Differential Revision: D69073908