Make Module:backward() final

[jason-dai]

We should make Module:backward final, and overload the updategradInput and accGradParameters methods in its subclasses instead.

[zhichao-li]

A similar one need to be final: #966

[yiheng]

Forward and backward should be final for all non-container layers. The non-container layer should only override updateOutput(), updateGradInput() and accGrad(). Some common logic(e.g. bypass the layer/freeze the parameter) should be put into the forward/backward method and not be changed by the non-container layers.

Forward/Backward should be overridden in container layers. In container forward/backward method, the submodules should always call forward/backward method of the submodule.