chore: add tests for ash_sql #243

[matt-beanland]

Contributor checklist

Leave anything that you believe does not apply unchecked.

• I accept the AI Policy, or AI was not used in the creation of this PR.
• Bug fixes include regression tests
• Chores
• Documentation changes
• Features include unit/acceptance tests
• Refactoring
• Update dependencies

Reproduction/regression test for: ash-project/ash_sql#244 (issue:
ash-project/ash_sql#243).

Ash.Query.combination_of/2 applies its parts in the order given, each to the result of
everything before it. ash_sql renders them as a flat chain of set operations, so SQL's own
precedence regroups a combination that spans more than one precedence level .
INTERSECTbinds tighter thanUNION/EXCEPT. The query then answers differently here than it does under Ash.DataLayer.Ets`.

The test adds a three-part case to test/combination_test.exs: base, union, then intersect,
where applying the parts in order and letting SQL regroup them give different answers. A fourth
record belongs to neither grouping, so reading the whole table cannot be mistaken for a pass.

against ash_sql 0.6.6      left:  ["alpha", "beta", "gamma"]
                           right: ["alpha", "gamma"]

with ash_sql#NNN           19 passed

Every existing combination test here stops at two parts, which is the one shape where the order
cannot be observed — two parts agree however they are grouped, and union/except share a
precedence in SQL so they agree too. That is why this has not surfaced.

⚠️ This test fails until ash_sql#244 lands and is released. Note also that main is currently red
independently of this: 11 tests in AshPostgres.AggregateReadActionTest fail on fc165f3e
against released ash_sql 0.6.6, unrelated to combinations or to this change.

[zachdaniel]

🚀 Thank you for your contribution! 🚀