Infinite Retrieval: Attention Enhanced LLMs in Long-Context Processing