Turn feature combining in TED/DIET into re-usable layers

[samsucik]

We want to simplify the code of TED and DIET by factoring out the combining of different feature types into re-usable model components (layers).

The implementation of TransformerRasaModel._combine_sparse_dense_features and TransformerRasaModel._combine_sequence_sentence_features will be made into re-usable layers inheriting from tf.keras.layers.Layer. This way, the layers can be used more generally, not just in Transformer models, and the architecture implementation of TEDPolicy and DIETClassifier will be simplified, better reflecting the high-level schemata of the models' architectures.

Actual work happens in #7589.