Speculative Speculative Decoding For Faster Llm Inference