Created by: reyoung

The current backward algorithm in backward.cc has several problems:

1. There are two kinds of operators, computational and control flow. The backward logic of these two kinds could be significantly different.
2. We should make users be able to customize while backwards a neural network.
   • The error clipping will add crop operators when backward.
3. The variable name is not the only key of the gradient in no_grad_set. It should be a pair of variable name and its block id.
4. We should propagate no_grad after renaming the variable.