Efficient Memory Management for Large Language Model Serving with PagedAttention