Rank studies using new online learning framework

[bdewilde]

changes

• adds a new StudyRanker class with underlying featurizer+classifier model, that works generically for both citations and fulltexts, in either online or batch training contexts
• adds new tasks to train study ranker model in online/batch contexts, depending on inputs; removes old tasks to train citation ranker model
• simplifies and extends situations that result in training of the study ranker
• applies study ranker model in relevance ranking of studies
• adds river dependency, and bumps numpy, used in new model