[FEA] Ability to export cuML RF models and run prediction on machines without a GPU

[hcho3]

Is your feature request related to a problem? Please describe.
This is similar to #3556 and #3822, but for random forests specifically.

Describe the solution you'd like
When the latest Treelite (1.3.0) is brought into cuML, we will be able to convert cuML RF to a Treelite object and then serialize it as a checkpoint. Treelite now offers a binary checkpoint format so that tree models can be exchanged between different machines.

Running prediction on machines without GPUs will proceed as follows:

1. Export cuML RF as a Treelite checkpoint checkpoint.tl.
2. Copy the checkpoint file checkpoint.tl to the target machine (which has no GPU)
3. Install Treelite package on the target machine.
4. Load the checkpoint checkpoint.tl on the target machine, by calling treelite.Model.deserialize(...).
5. Make prediction, by calling treelite.gtil.predict(...).

Compatibility considerations.

Checkpoints are specific to the version of Treelite with which it was produced. Each version of cuML is pinned to a specific version of Treelite, and the target machine must have the exact version of Treelite installed.

For example, the upcoming version of cuML will have Treelite 1.3.0, so the target machine must also have Treelite 1.3.0 installed.

[hcho3]

Closing, since it's now possible to export cuML RF models as a checkpoint file that can be later be loaded into a machine without a GPU.

from cuml.ensemble import RandomForestClassifier as cumlRandomForestClassifier
from sklearn.datasets import load_iris

X, y = load_iris(return_X_y=True)
X, y = X.astype(np.float32), y.astype(np.int32)
clf = cumlRandomForestClassifier(max_depth=3, random_state=0, n_estimators=10)
clf.fit(X, y)

checkpoint_path = './checkpoint.tl'
# Export cuML RF model as Treelite checkpoint
clf.convert_to_treelite_model().to_treelite_checkpoint(checkpoint_path)

Later on a machine without a GPU:

import treelite

# The checkpoint file has been copied over
checkpoint_path = './checkpoint.tl'
tl_model = treelite.Model.deserialize(checkpoint_path)
out_prob = treelite.gtil.predict(tl_model, X)

Note that only Treelite needs to be installed on the target machine; cuML is not required.