Several minor bugfixes

[desilinguist]

• Allow non-default base estimators for AdaBoost (Base estimators other than default are not supported for AdaBoost #238)
  • Correctly interpret the base_estimator fixed parameter for AdaBoostClassifiers and AdaBoostRegressors as objects instead of strings.
  • Make sure that the base estimator for Adaboost has a fixed seed so that results are replicable.
  • Add corresponding unit tests for both classification and regression and update documentation.
• Fix hasher_features typo in experiments.py (Bug when checking if hasher_features is a valid option #234).
• Remove unnecessary setting of numpy global random seeds (Stop modifying global numpy random seed #220)
  • We are already passing in random_state seeds in model_kwargs for all classifiers/regressors so we don't need to set the numpy global random seed.
  • Include Lasso, ElasticNet, LinearSVR, and SVC when setting the above seeds in model_kwargs.
• Remove requirements for several learners to have dense input (Remove decision tree and random forest learners from requires_dense set #207)
  • In the new version of scikit-learn, only GradientBoostingClassifier and GradientBoostingRegressor require dense input.
  • Move make_sparse_data() to utils.py
  • Update test_sparse_predict() to include all sparse-friendly classifiers.
• Update expected values in certain tests to account for all of the above changes.

[dan-blanchard]

👍 I looked at this on my phone at lunch and forgot to comment. Looks good to me.

[aoifecahill]

Looks good to me too, thanks!