BUG: Fix aggregate exploding the output of Reduction ops that return a list/ndarray

[timothydijamco]

Problem

On the Pandas backend, when .aggregate() is passed an aggregation/reduction that results in a list/ndarray, aggregate will automatically explode that list/ndarray into many rows (one row per element in the array). This is not the desired behavior. A list/ndarray coming from a Reduction op should actually be treated as a scalar (i.e., should result in a single row, not many rows).

Example code

## Setup Pandas backend
df = pd.DataFrame({
    'key': pd.Series([1, 1, 2]),
    'foo': pd.Series([100, 200, 300]),
})

ibis_pandas = ibis.pandas.connect({'test': df})

t = ibis_pandas.table('test')


## Create two Reduction UDFs: one that returns a list, and one that returns an ndarray
@udf.reduction(input_type=[dt.int32], output_type=dt.Array(dt.int32))
def to_list_udf(v):
    return list(v)

@udf.reduction(input_type=[dt.int32], output_type=dt.Array(dt.int32))
def to_np_array_udf(v):
    return np.array(v)


## Use UDFs in `aggregate`
ibis_pandas.execute(
    t.aggregate(
        foo_list=to_list_udf(t['foo']),
        foo_np_array=to_np_array_udf(t['foo']),
    )
)

Current output

	foo_list	foo_np_array
0	100	100
1	200	200
2	300	300

Desired output

	foo_list	foo_np_array
0	[100, 200, 300]	[100 200 300]

Solution

This PR fixes the coerce_to_output function used by .aggregate() so that it treats lists or ndarrays from Reduction ops as scalars, and doesn't explode them into multiple rows.

[jreback]

looks fine, can you add this pr number to the original whatsnew note. ping on green.

[timothydijamco]

@jreback added a release note, and CI is green!

[jreback]

thanks @timothydijamco