Brain-inspired sparse training enables Transformers and LLMs to perform as fully connected