Unexpected behavior when using drop_nans() on all columns in Polars