Expose hyperparameters property for sklearn classifiers + stream.train fix for quiver-based classifiers

In SklearnModel, I added a hyperparameters property to expose a dictionary of hyperparameters used in the classifier. These same hyperparameters are now printed if verbose = True in the training jobs.

While testing this change, I found a bug in the streaming training jobs where new_umbrella is called for the initial lookback jobs which doesn't exist, only the umbrella exists at this point, so I pointed the search for target, random times to point at umbrella instead. Not sure how this came about, but the training jobs appear to work now.