refactor: unify default array representation in Pandas and Dask backends

#2619 will cover making the Pandas and Dask backends accept either np.arrays or lists (in other words, be able to execute operations on array expressions, regardless of whether they are represented by lists or by np.arrays).

However, the Ibis API has several functions that can create array expressions for the user, and unfortunately, after #2615, the Pandas and Dask backends are no longer consistent in what kind of arrays they create—they may output either lists or np.arrays, depending on the part of the API.

API	Operation	Creates	Represeted by
ibis.literal([1, 2, 3])	Literal	ArrayScalar	list
ibis.literal(np.array([1, 2, 3]))	Literal	ArrayScalar	np.array
ibis.array([1, 2, 3])	Literal	ArrayScalar	list
ibis.array(np.array([1, 2, 3]))	Literal	ArrayScalar	list
ibis.array([t.int_col, t.other_int_col])	ArrayColumn	ArrayColumn	np.arrays (elements)
arr_scalar[1:3]	ArraySlice	ArrayScalar	list
t.arr_col[1:3]	ArraySlice	ArrayColumn	list (elements)
arr_scalar + other_arr_scalar	ArrayConcat	ArrayScalar	list
t.arr_col + arr_scalar	ArrayConcat	ArrayColumn	💥 (not working)
t.arr_col + t.other_arr_col	ArrayConcat	ArrayColumn	list (elements)
arr_scalar * 5	ArrayRepeat	ArrayScalar	list
t.arr_col * 5	ArrayRepeat	ArrayColumn	list (elements)
t.int_col.collect()	ArrayCollect	ArrayScalar	list
t.int_col.quantile([0.25, 0.50, 0.75])	MultiQuantile	ArrayScalar	list
map_scalar.values()	MapValues	ArrayScalar	list
t.map_col.values()	MapValues	ArrayColumn	list (elements)
map_scalar.keys()	MapKeys	ArrayScalar	list
t.map_col.keys()	MapKeys	ArrayColumn	list (elements)
t.str_col.split(',')	StringSplit	ArrayColumn	list (elements)
elem_udf_list(t.int_col)	ElementWiseVectorizedUDF	ArrayColumn	list (elements)
elem_udf_ndarray(t.int_col)	ElementWiseVectorizedUDF	ArrayColumn	np.arrays (elements)
analytic_udf_list(t.int_col)	AnalyticVectorizedUDF	ArrayColumn	list (elements)
analytic_udf_ndarray(t.int_col)	AnalyticVectorizedUDF	ArrayColumn	np.arrays (elements)
reduc_udf_list(t.int_col)	ReductionVectorizedUDF	ArrayScalar	list

We should make the return values for all of these operations consistent (otherwise it’s hard the user to predict what their arrays in the final executed DataFrame will look like), but I wanted to get some feedback on whether we want the default representation to be list or np.array.

I think arrays as np.arrays would be useful and I don’t see much reason to prefer lists over np.arrays—however I think most other backends represent arrays using lists (in the Pandas DataFrames resulting from execution).

[tswast]

Related: https://github.com/ibis-project/ibis/issues/2377 has some thoughts on supporting lists of structs via Arrow.

[cpcloud]

@timothydijamco Thanks for the great issue write up. It really warms my heart to get issues with descriptions like this. It makes maintainers lives really easy.