* Collect output scale for quantized op and fused op
* Post_training_quantizaion sets batch_generator to support lod tensor