Managing nullable types and empty partitions

Hi, I'd like to be able to handle nullable types and empty partitions before making a prediction in Dask-ML.

With the former, Dask DataFrame columns with nullable types can be recast to non-nullable types. With the latter, the empty partitions of the Dask DataFrame can be dropped.

Start at the Dask-ML prediction entry point that receives a Dask DataFrame and inspect how nullable columns and empty partitions are handled before prediction. Add focused tests for nullable-to-non-nullable conversion and empty-partition handling, then verify prediction succeeds for both cases.

Written by the indexing model from the issue text.