Fix quantize kernel while pad != 0, implement the fusion kernel of dequantize, elementwise_add, batch norm and Relu !1316
Created by: hjchen2
quantize kernel支持padding,使用模型转换工具 https://github.com/hjchen2/paddle-test/blob/master/convert_caffe2fluid/convert_to_quantize.py 可以将int8模型中conv的padding前置到quantize op中。
padding前置目前只对8bit mobilenet-v1进行过测试,预测速度提升10ms+。