Add support for dynamic real features to tempflow

[NielsRogge]

This PR:

• removes cardinality and embedding_dimension from TransformerTempFlow, because currently the model is initialized to get it, although it doesn't use it. So removing these avoids confusion.
• adds support for dynamic real features, such as holiday information.

Is it correct that static features are not useful for multivariate models like TransformerTempFlow and TimeGrad (as there is only a single time series, and no distinction is made between individual time series)? Or can these models still learn feature embeddings for each individual time series?

[kashif]

@NielsRogge thanks! So yes regarding your questions... if there would be static features (i.e. when you have many multivariate time series and want to train a global model on those then yes features could help). So yes I think there is no point in having say embedding for individual time series... it will just not change those embeddings and just bloat the model.