Created by: slf12

add power to quant_embedding dict 将dequantize_log时的运算power挪在量化压缩时，以加速模型inference速度。对应paddle pr https://github.com/PaddlePaddle/Paddle/pull/24607