The N-Grammys: Accelerating Autoregressive Inference with Learning-Free Batched Speculation