Simplify DNN dense corpus and interaction with python backend

[johann-petrak]

• see also IMPORTANT! Refactor and simplify gate-lf-python-data#15
• Keep the option to have many features but make it easy to have just the simple one-feature approach.
• Store dense corpus instances as maps in each line with standard keys for label and possibly features
• unify with representation for unlabeled data (e.g. embedding creation or topic models) and other kinds of supervised/unsupervised tasks, e.g. seq2seq or semantic similarity
• !!!! change representation of sequences: instead of having a sequence of element with multiple features, have a sequence for each feature. Makes it MUCH easier to create batches later.
• Make it easy to swith between our output and the torchnlp library in the python backend

This should become a project possibly with several subissues.