Keep runnable tasks queued across revisions

[Mazyod]

Fixes moby/moby#53722

Related to moby/moby#36699. Its later node-restart reports may share this cause, but lack the service-revision evidence needed to confirm that. The original issue concerns obsolete tasks desired Shutdown, whose behavior this patch preserves.

- What I did

Fix a valid Pending task being dropped from scheduler retries after a service-only metadata update. If no node can accept the task during a subsequent scheduler pass, the older-revision branch currently skips re-enqueueing it even though it is still desired Running. It can remain Pending after node recovery until a force update creates a replacement.

- How I did it

Require desired state Shutdown or later before taking the older-revision retirement branch in noSuitableNode. Still-runnable tasks reach the existing re-enqueue path. Add one regression with identical task specifications and different revision numbers, checking queue retention and assignment of the same task ID after Down becomes Ready.

- How to test it

Base: 1fd637ba5cc32ff30d1dd2bdb14997bd4f424b46; Go 1.27.1 linux/amd64.

• Regression only on stock: go test ./manager/scheduler -run '^TestPendingOldRevisionRecoversWhenNodeReturns$' -count=1 fails because the task is absent from unassignedTasks.
• With the fix: go test ./manager/scheduler -count=1 and go test -race ./manager/scheduler -count=1 pass. The existing TestUnscheduleableTask continues to verify retirement of an obsolete Pending task.
• go test -parallel 8 -timeout 20m ./... -count=1 passes for the full root module, including integration, jobs, task reaper, and updater packages. A short temporary path is needed for the node tests' Unix sockets in a deeply nested checkout.
• Fresh stock and patched Engine builds from Moby 29.8.1 were compared in disposable Linux amd64 DinD. Stock reproduces both pause/activate and worker stop/start; the patched manager recovers the same Pending task IDs in both, including with a stock worker. Only the identical vendored scheduler file was changed.

Runtime comparison: Moby 464cd50c3d9e92877d56940ea160de6fca7bea23, vendoring SwarmKit ad0357aedfca72c288abb8395ec5603e1136a9e5; Go 1.27.1, CGO_ENABLED=0, tags netgo osusergo static_build nri_no_wasm. Binary SHA-256: stock 34747d599445c61b0a79e1e7dc7bd175ce9f3a9885576cded4a6d16d7b46e6c6; fixed d2fd714b5dcddc237192cff9a1da44774e0874209da1b81593dc475cafecf495. The separate swarmd module and full Moby suite were not run.

- Description for the changelog

Keep valid pending tasks eligible for scheduling after service metadata updates and node recovery.

Created with: Astra (AI), under my supervision and review.

[koriyoshi2041]

Verified the scheduler ownership boundary at f54226a: only older-revision tasks already desired Shutdown or later take the retirement path; runnable and job-completion targets stay on the existing retry path. The regression also preserves the original task ID across node recovery, while TestUnscheduleableTask continues to cover obsolete Pending-task shutdown. Full scheduler tests and the race-enabled scheduler suite pass locally; diff hygiene is clean.