* add new model, bmn dygraph model for video temporal proposal generation

* matching new API, decoupling model and  train module