Separate larger submodules from layers

[oskar-taubert]

Currently the package seems to be organized in models (e.g. Transformer) and modules (e.g. attention) so anything that is not a complete model.
It would make sense to me to add a tier in the middle so there would be

• Models (e.g. Transformer)
• Components/Modules/Blocks (e.g. Encoder, Decoder, OutputModule, etc.)
• Layers (MCDropout, attention, etc.)