Refactor LLM runner to bypass the upstream ET runner

Currently the LLMRunner is a thin wrapper over ET's native runner which is experimental and doesn't support many features that it should (e.g. batched prefill - required for models exported with seqlen < ctxlen, custom sampling strategies, constrained decoding, etc.). The scope of this issue is to reimplement it inside TS using the Model primitive and drive the generation directly from TS.