Decentralize development with layer parameter registration mechanism

[futurely]

To get rid of the last obstacle in development (#3315 (comment)), one method proven effective is MXNet's layer registration and custom model serialization. Different users can plug in their own layer implementations without static centralized coordination since the model definitions and parameters are all saved to and loaded from dynamically generated JSON and binary stream.

If caffe.proto can also be dynamically generated from all the layer headers, then the users will still be able to use pure text files to define networks and develop without conflicts at the same time.

[futurely]

TensorFlow uses a similar Op (layer) registration mechanism.

[ih4cku]

That would be great, look forward to see.