bugfix. gradient of y_c w.r.t feature_map

[insikk]

• Use y_c (logit before softmax) for calculating target conv layer gradient
• Remove unnecessary gradient normalization for target conv layer gradient
• Use multi-class cross entropy loss for guided backpropagation
• For VGG network, we use train_mode=False always to ensure dropout is turned off
• For Resnet network, we use train_mode=True, which is default value, always to ensure consistency
  • There is issue when we pass train_mode=False, inference is ruined due to bad design on batch-norm layer (see https://github.com/tensorflow/tensorflow/issues/4887)
• On summation of weighted feature map, we use zero-filled array for initial value, not one-filled array.