Accelerating Large Language Model Decoding with Speculative Sampling