[ENH]Warn users when training non dask friendly ML models with wrap_predict=False.

[VibhuJawa]

Is your feature request related to a problem? Please describe.

We should warn users when training non dask friendly ML models with wrap_predict=False. This is because when wrap_predict=False we assume that the model can handle non dask objects.

https://github.com/dask-contrib/dask-sql/blob/98c38cfba0ca3a08f539523c6480cc2736b505cd/dask_sql/physical/rel/custom/create_model.py#L189-L190

This is often not true for various model classes like xgboost.XGBRegressor

Describe the solution you'd like We should create loud warnings based on the model class as the models that can handle dask often have dask so if we dont detect that we should potentially wan users.

CC: @randerzander