return_all_estimators option in GridSearch and RandomizedSearch

[Thomasillo]

Describe the workflow you want to enable

I would like to have an option in the GridSearchCV and RandomizedSearchCV
that enables storing of the (re)fitted estimator for all parameter combinations.

The background is that I'm working with highly correlated timeseries data and try to find the best
way to set up cross validation.

I have a test set set aside and want to examine for which cross validation strategy, the CV score correlates best with the score on the test set.

Currently, only the "best" estimator is stored thus not allowing to evaluate the others on the test set.

Describe your proposed solution

A slight modification of BaseSearchCV should do the trick. I would give it a shot if appreciated and the above makes sense.

[raduspaimoc]

@Thomasillo Hey, Is there any way in which I could contribute to this new feature?

[chrispe]

I would also be interested to work on this. Has it been picked up by somebody else already? Thanks.

[chrispe]

Since nobody responded, I took it on and submitted a pull request as you see above.