NEZHA: A Zero-sacrifice and Hyperspeed Decoding Architecture for Generative Recommendations