SUBLLM: A Novel Efficient Architecture with Token Sequence Subsampling for LLM