FIX/ENH: make explicit scoring for 1st level estimators

[dengemann]

Allow to set globally or specifically for estimators the scoring function to be used. Currently which one is used somehow fuzzily depends on whether an estimator has a predict proba or decision function.

cc @mrahim