CI: add Cloudberry main support - #124

CI was pinned to Cloudberry REL_2_STABLE in deeb094 because main moved
to the PG16 kernel and broke some tests. That left Cloudberry main
(the future 3.x) without any coverage.

Turn the Cloudberry branch into a matrix dimension, so that every test
target runs against both REL_2_STABLE and main. Artifact names and job
names now include the branch. The test job runs with !cancelled(), so a
failed build on one branch does not skip the test jobs of the other.

Also fix the regression target on main. Its minimal_schedule tests now
expect the shared tables (char_tbl, int2_tbl, ...) to exist already,
and those tables are created by test_setup. Without test_setup, 5 of
the 7 tests failed and the regression database being backed up was
almost empty. Run test_setup first wherever it exists. On
REL_2_STABLE all tests still pass and the schema diff is unchanged.

Tested by running every target on a demo cluster on Cloudberry main
in the CI build image.

Assisted-by: Claude Code

GetForeignDataWrappers() and GetForeignServers() have no ORDER BY, so
the database may return their rows in any order. The integration test
"creates foreign data wrappers with a validator and options" indexes
the result as [0] and [1], and it failed on the Cloudberry main leg of
CI with the two wrappers swapped. The same run had passed earlier on a
local main cluster only by chance.

Order both queries by oid. That makes the tests deterministic, and the
backup now writes these objects in a stable order too.

This change is taken from Jinbao Chen's PR apache#106.

See: apache#106

Co-authored-by: Jinbao Chen <chenjinbao1989@gmail.com>
Assisted-by: Claude Code