pystruct: use rankSVM features as unary features

[chengsoonong]

Removes training of rankSVM, and gives more interpretable weights.

[cdawei]

It seems to be nontrivial to remove training of rankSVM, as it provides a way to generate node features when prediction, i.e., duplicate the ranking probabilities of all POIs (traj_length - 2) times. If rankSVM was not trained, for prediction, when only (start, end, length) are known, there needs a way to generate node features.

[lexingxie]

this indeed non-trival.

Therefore as suggested in our meeting yesterday -- keep the current implementation for now (until the recsys deadline), and then implement/resolve this issue after the submission deadline.

[cdawei]

Merge with issue #81