feat(api): support order_by in order-sensitive aggregates (collect/group_concat/first/last)

[jcrist]

This adds a new order_by keyword argument to order-sensitive aggregates (collect/group_concat/first/last) to specify the ordering the aggregation should use. By default no ordering is specified, and the result is backend dependent (following the current behavior).

In [1]: import ibis

In [2]: t = ibis.memtable({"x": [1, 3, 2, 4], "y": [10, 9, 8, 7]}, name="test")

In [3]: t.x.collect().execute()  # no ordering specified, backend-dependent result
Out[3]: [1, 3, 2, 4]

In [4]: t.x.collect(order_by="y").execute()  # use the specified ordering, or error if unsupported
Out[4]: [4, 2, 3, 1]

In [5]: ibis.to_sql(t.x.collect(order_by="y"))
Out[5]: 
SELECT
  ARRAY_AGG("t0"."x" ORDER BY "t0"."y" ASC) FILTER(WHERE
    "t0"."x" IS NOT NULL) AS "ArrayCollect(x, (y,))"
FROM "test" AS "t0"

This is generally implemented using either an aggregate-internal ORDER BY clause, or via WITHIN GROUP (...) (backend dependent). The SQL required here can vary significantly between backends unfortunately :/.

Fixes #9170.

[jcrist]

I don't have cloud testing setup locally, so I'm 🤞 that I did everything correctly for bigquery/snowflake.

[cpcloud]

Thanks!

One concern about ignoring filters with literal inputs.

[cpcloud]

There's a couple things that can be pulled out here, I'll do a couple separate PRs.

1. RisingWave seems to support UNNEST now in test(risingwave): mark test_union_aliasing as notimpl because arbitrary aggregate function is not supported #9731.
2. Oracle group_concat was just implemented incorrectly, but is pretty easy to get working in feat(oracle): support group_concat operator #9732.

[cpcloud]

I think there's probably an argument to be made for adding order_by to arbitrary but that can be done in a follow up.

[cpcloud]

Sweet, thanks @jcrist! Merging.