Support neural network models that have layers connected in a directed acyclic graph (DAG)

[beomyeol]

In the current implementation, the layers in a neural network can be connected in a linear fashion, which means that one layer is only connected to its previous layer and its next layer. However, the state-of-the-art CNN models such as GoogleNet, layers are connected in computation graph - a directed acyclic graph. Thus, like other deep neural network frameworks such as Caffe, dolphin-dnn should train neural network models that consist of layers connected in a DAG.

This issue is brought from cmssnu/dolphin#153.