[ENH] Backend agnostic machine learning models

[VibhuJawa]

Is your feature request related to a problem? Please describe. As we are working becoming more backend agnostic (GPU/CPU) , we should look into a way of supporting multiple ML backends with minimal code change .

Describe the solution you'd like

We currently have to specify the cuml/sklearn model class. We should look into a way of training models where we detect what the input type is and have the user just specify ‘RandomForest’ and have dask-sql handle inferring the rest of the classname.

So if the training dataframe is on CPU we use sklearn and if it is on GPU we use cuML.