separate scoring from train/test

[kegl]

At this time y_pred objects are wrapped into Predictions types, read out from there, and passed to scoring. Two issues: 1) they are not converted to and deconverted from zipped numpy arrays, and 2) scores are not run independently, so if someone created common objects in score, they work with ramp_test_submission but not on the server.

We could keep the second feature, make sure it works in the same way in ramp-engine. But for the first, we should separate scoring from train test, and test that scoring works on saved/loaded predictions. There is a default now, slightly different here and in board, we should make sure that the conversion-deconversion code is the same. I think save/load or convert/deconvert should be implemented as a member class in Predictions, so if someone wants a custom code, they would have to implement a new derived Predictions class.