Speed Always Wins: A Survey on Efficient Architectures for LLMs