OnDemandFeatureView: Allow python transformations for offline processing.

[job-almekinders]

Is your feature request related to a problem? Please describe.
When running a PythonTransformation for offline feature retrieval, an exception is raised here. I understand that python transformations on very large datasets do not scale well. However, in some cases, a user might still want to apply a python transformation to a smaller dataset in the offline store. Due to this error (and the missing implementation of PythonTransformation().transform_arrow()), this is currently not possible.

Describe the solution you'd like
Instead of raising an exception, we could still run the transformation, and also log a warning indicating that it may not scale well. On top of that we can implement the functionality in transform_arrow to make the transformation happen.

Describe alternatives you've considered

Additional context

[franciscojavierarceo]

Oh we should definitely discuss this because I agree. Also, for scaling Spark should really call the Transformation as a UDF and then it can scale very well.

[franciscojavierarceo]

This has been implemented. 👍

[job-almekinders]

Thanks a lot for this! @franciscojavierarceo

FYI @TomSteenbergen :)