Article 2 Speculative Decoding in Practice: How EAGLE3 Makes LLMs Faster Without Changing Their Outputs