* add new op paddle.tensor.div(x, y, out=None, name=None) 
* add gpu and dygraph unittests.
* Performance optimization: scale op is not called when alpha=1. 
* op error message optimization.