Text this: Attention Once Is All You Need: Efficient Streaming Inference with Stateful Transformers