Created by: juncaipeng

Optimize obtaining quantized scale in inference stage

为了减小加载模型的耗时，量化op构建中直接使用保存scale的名字，从opdesc中拿到scale。