Scatter update implemented with in-place memory, gradient_check modified a little bit to suit in-place !3668
Created by: zchen0211
I implemented this function. https://www.tensorflow.org/api_docs/python/tf/scatter_update with the add case. Use in-place memory for forward and backward, with Python unit-test passed. Also, I modified the gradient check a little bit (reset data back to origin) to make sure the numerical gradient is computed correctly.