Inevitable Trade-off between Watermark Strength and Speculative Sampling Efficiency for Language Models