map_partitions doesn't work with Scalars

dask/dask works here, not sure how important this is though.

reproducer:

```
import pandas as pd

pdf = pd.DataFrame({"a": [1, 2]})
df = from_pandas(npartitions=2)

map_partitions(lambda x: x, df.a.sum())
```

Start by running the pandas reproducer involving from_pandas, df.a.sum(), and map_partitions(lambda x: x, ...). Compare its behavior with dask/dask, as the issue reports that implementation works there. Done means the scalar case behaves consistently, with a regression test covering the reproducer.

Written by the indexing model from the issue text.