SetTransformer and other Summary Network Issues

[LarsKue]

Summary networks in general are still poorly tested. Upon introducing more rigorous testing, I encountered some issues.

Set Transformer

@stefanradev93
The set transformer (along with its child classes like MAB, PMA, etc.) appears to have multiple issues:

1. It cannot build automatically, the user has to supply the input dimension.
2. Unlike other networks, it does not properly handle converting user-friendly inputs like dropout=None or dropout=0.
3. Its argument naming conventions differ from the other networks, e.g. num_attention_blocks instead of depth.
4. It relies on manually named argument forwarding, e.g. num_dense_feedforward cannot easily be renamed without sweeping the whole directory.

How do you suggest we move forward with this? Upon writing more rigorous tests, I find (1) particularly bothersome.

(to be expanded)

[stefanradev93]

The set transformer hasn't been touched for a few months now. I will fix 1 and 2. We need to talk about 3 and 4 (e.g., depth being a too generic argument for a composite network).