Make rebalance colocation-group representative selection deterministic

[ibrahim halatci (ihalatci)]

This is a product defect, not a test flake. It was found via a nightly background_rebalance_parallel_reference_tables
failure, but the test was reporting the bug correctly.

Root cause

NonColocatedDistRelationIdList() (shard_rebalancer.c:2013-2058) builds its list from
CitusTableTypeIdList() and then keeps the first table seen per colocationId (the dedup
continue at :2050).

CitusTableTypeIdList() scans pg_dist_partition with an unordered systable_beginscan
(metadata_cache.c:5078-5080). So the representative table for each colocation group is whichever
tuple physical order happens to yield — it can change after a HOT update, a vacuum, or any other
storage-level churn, with no schema or data change at all.

Two consequences:

1. The representative relation name is printed in DEBUG1 output.
2. The shard IDs emitted in the rebalance plan are those of whichever table won.

The second is the serious one, and it also exposes the sibling test background_rebalance_parallel,
which is only shielded from the first because it never raises client_min_messages to DEBUG1. That
ruled out a test-only fix.

Reproduced on demand

Two arms on a fresh 3-node cluster with a forced HOT update on pg_dist_partition, differing only in
the installed library (verified binary-identical to the intended build at both ends of each arm):

Library	representative before churn	after churn
Unpatched	table1_colg1	table2_colg1 ← bug reproduces
Patched	table1_colg1	table1_colg1 ← fixed

Measured OIDs: 17320 table1_colg1 < 17326 table2_colg1.

Fix

Sort by OID before the de-duplication. Lowest OID is creation order, which is exactly what the
existing committed baselines already assert — so this converts an incidental property into a
guaranteed one rather than changing behaviour.

The sort is in-place on a list built fresh per call, and needs no new #include.

Validation

• Full check-operations — 17/17 pass, with the installed library MD5-verified against the build
  tree both before and after the run.
• Zero expected-output churn. git status --porcelain -- src/test/regress/ is empty; no .out and
  no .sql file is touched by this PR. That is the actual evidence this is a real fix and not
  test-shaping.
• Six-call-site audit of NonColocatedDistRelationIdList() consumers (:1138, :1179, :1334,
  :1445, :3787, :4051) — all clean; none depends on the previous arbitrary order.

Deliberately not changed

CitusTableTypeIdList() in metadata_cache.c is left unordered. Sorting at the source would impose
cost on every caller; the rebalancer is the one that needs the ordering guarantee.

Refs #8776

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 88.68%. Comparing base (52663ed) to head (aa99c96).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #8789      +/-   ##
==========================================
- Coverage   88.70%   88.68%   -0.02%     
==========================================
  Files         289      289              
  Lines       64889    64890       +1     
  Branches     8181     8180       -1     
==========================================
- Hits        57559    57548      -11     
- Misses       4967     4976       +9     
- Partials     2363     2366       +3     

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.