• Work with result dicts rather than tuples from fitting each fold, being @ogrisel's proposal for the native storage of search results (with the intention to produce a fault-tolerant log of all fold results.)
• Refactor grid_search and cross_validation.cross_val_score into a CVScorer, which I think should be private for this release until we're happy with its interface.

This will make it much easier to support returning training times, etc. and additional input such as sample_weight.

One of the main problem with dicts is that, unlike attributes of objects, the field names are difficult to deprecate (though if we needed to, we might consider subclassing dict).

It is part of the way towards a more extensible results interface from the *SearchCV estimators, as discussed at the ML, #1787, #1768, and to some extent #1034, #1020.

Note: the CVScorer introduced in this PR is missing tests as yet.