pipeline: separate the model define plugin into 2 lifecycle

As for the current __model-define__ plugin, we use the one function to handle the different lifecycles: training and serving, this looks a little complicated to plugin developers. Therefore we want to address the issue by the following:

1. add an optional new file "inference.js" in every __model-define__ plugin, which will be called when serving.
2. pass the labelmap to "inference.js".