opt: InlineWith should not inline mutations into subqueries and apply-joins

[mgartner]

The InlineWith normalization rule can inline a mutation into subqueries or apply-joins, causing the mutations to be executed more than once. For example, notice that the INSERT is executed five times:

defaultdb> SELECT version();
                                          version
-------------------------------------------------------------------------------------------
  CockroachDB CCL v22.2.2 (aarch64-apple-darwin21.2, built 2023/01/04 17:38:41, go1.19.1)
(1 row)

defaultdb> CREATE TABLE corr ( k INT PRIMARY KEY, i INT);
CREATE TABLE

defaultdb> CREATE TABLE corr2 (k INT, i INT);
CREATE TABLE

defaultdb> INSERT INTO corr VALUES (1, 10), (2, 22), (3, 30), (4, 40), (5, 50);
INSERT 0 5

defaultdb> WITH insVals AS NOT MATERIALIZED (INSERT INTO corr2 VALUES (1, 10) RETURNING k)
SELECT * FROM corr
WHERE EXISTS (SELECT * FROM insVals WHERE k = corr.k UNION ALL SELECT k FROM corr c3 WHERE k = corr.k);
  k | i
----+-----
  1 | 10
  2 | 22
  3 | 30
  4 | 40
  5 | 50
(5 rows)

defaultdb> SELECT * FROM corr2;
  k | i
----+-----
  1 | 10
  1 | 10
  1 | 10
  1 | 10
  1 | 10
(5 rows)

This is not consistent with Postgres which guarantees that:

Data-modifying statements in WITH are executed exactly once, and always to completion, independently of whether the primary query reads all (or indeed any) of their output.

^ From the Postgres Docs.

We should prevent InlineWith from inlining into correlated subqueries (and maybe uncorrelated, unless they are guaranteed to execute onle once - even lazy ones?), and apply-joins.

Jira issue: CRDB-23481

[mgartner]

The PG docs don't mention specifically what they behavior is for data-modifying statements in WITHs that have NOT MATERIALIZED, but the answer is in the source code:

https://github.com/postgres/postgres/blob/15eb1228800a35042ef318f941173aa8b89002d1/src/backend/optimizer/plan/subselect.c#L952-L957

NOT MATERIALIZED (or CTEMaterializeNever) is ignored if the statement inside the WITH is a DML statement, like INSERT.