Deprecate predict_model in cuml.ensemble/cuml.dask.ensemble

[jcrist]

The predict methods of our ensemble estimators (cuml.ensemble.RandomForestClassifier, cuml.ensemble.RandomForestRegressor, and the dask equivalents) include an old predict_model kwarg. This kwarg may be one of ("GPU", "CPU"), with "GPU" as the default.

On single node:

• predict_model="GPU": used FIL on GPU to do inference
• predict_model="CPU": used the internal decision tree representation used for fitting (so tied to the GPU), but does the predict on CPU. This is less performant and still doesn't let users do inference without a GPU.

On multi node (dask):

• predict_model="GPU": used FIL on GPU to do inference (broadcasting the complete model to every worker node)
• predict_model="CPU": used the internal decision tree repr from fitting, but with some very inefficient post processing (looping over the full length of the data! building up a python list of responses!). Using broadcast_data=True seems to better match the intent here (more efficient with small datasets), and is available with the default FIL codepath.

This PR:

• Deprecates predict_model everywhere. The FIL codepath is always taken. A warning is raised if predict_model is explicitly set, but the old codepaths no longer exist.
• Deletes the old codepaths. We could also delete some C++ used for inference, but I left that out for now.
• Updates the tests accordingly (including tests for the deprecated behavior).

No existing code should be broken by this (at most users will see a warning telling them the kwargs are going away), and all use cases should still be supported without the predict_model kwarg.

This PR is part of cleaning up the memory management and internals of our cuml.ensemble estimators.

[csadorf]

LGTM, but please address my comments and suggestions.

[jcrist]

/merge