Implement Fusion Summary Networks

[vpratz]

Multimodal inputs occur in many settings. They require a more complex structure of summary_variables than numpy arrays, as different inputs usually have different dimensionality, or even number of dimensions. They might also necessitate different summary networks, which are tailored to the specific summary. Manual implementation of this is already possible with the current API, but streamlining it would be great for downstream tasks. To achieve this, we need

• a transform to introduce the additional nesting in the adapter: (e.g., fuse(keys, into="summary_variables") which creates a dict with the provided keys and corresponding values inside the summary_variables entry).
• a FusionSummaryNetwork which takes a dict of keys and corresponding summary networks, as well as an additional network that combines their outputs to a chosen summary dimension.

Open questions:

• Do we want to allow for transforms in the second level dict, for example by adding an adapter argument to fuse, which operates on the inner variables?

[stefanradev93]

Addressed in #473