Speed Always Wins: A Survey on Efficient Architectures for Large Language Models