Align folding of skill embeddings with folding underlying feedforward models (fnn/bnn)

[hosseinfani]

TODO or NOT TODO?
Current way of injecting skill embeddings are not fold-based here:

OpeNTF/src/main.py

Line 141 in dee9130

skill_vecs = t2v.get_dense_vecs(teamsvecs, vectype='skill')

The ntf models know the fold inside ntf.learn(), but since

• the folding remove the team-member links, and embeddings are mainly used for skills (transfer-based), and
• t2v.get_dense_vecs returns the vecs from t2v.model in the last fold (after going through the earlier folds)

I think it's no big deal for the underlying fnn/bnn, all their folds use the skill embeddings based on the graph of the last fold.