fix(ci): stop the e2e teardown from failing a green suite

[localai-org-maint-bot]

What changed

Two independent CI defects that together turn a passing E2E suite into a red check that burns a 6 hour runner slot.

1. teardown-e2e races the container reaper. run-e2e-image starts the container with -d --rm. docker stop returns as soon as the container exits, but Docker reaps a --rm container asynchronously after that. The docker rmi localai-tests on the next line then loses the race and fails:

Ran 235 of 236 Specs in 484.299 seconds
SUCCESS! -- 235 Passed | 0 Failed | 0 Pending | 1 Skipped
Test Suite Passed
...
docker rmi localai-tests
Error response from daemon: conflict: unable to remove repository reference "localai-tests"
  (must force) - container c8fca721d485 is using its referenced image 8be9d6846e5c
make: *** [Makefile:392: test-e2e] Error 1

Every spec had already passed. docker rm -f is synchronous, so the image reference is gone before teardown-e2e returns. The same change fixes a second latent bug on that line: with no matching container, docker stop $(docker ps -q ...) gets zero arguments and exits 1 with "requires at least 1 argument".

2. The tmate step holds the runner for 6 hours. It runs on every failure and, despite detached: true, goes silent until GitHub cancels the job at the wall. It now needs a ci-debug label on the pull request, with a 30 minute step timeout as a backstop. Applied to all five active workflows that carry the step, since the trap is identical in each.

Why

Sampled runs of tests-e2e.yml that hung and were cancelled at the 6 hour limit:

Run	Branch	Suite result	Cause of the red X
33435319093	bump/BONSAI_VERSION	235 passed, 0 failed	docker rmi conflict
33435332991	bump/DS4_VERSION	235 passed, 0 failed	docker rmi conflict
33412165884	master	235 passed, 0 failed	docker rmi conflict
33444669207	master	235 passed, 0 failed	docker rmi conflict
33438683472	test/distributed-e2e-ci	6 failed	genuine spec failure

The run list for this workflow currently shows repeated 6h, 7h, 9h and 12h cancelled runs.

How to verify

The race reproduces against a container shaped like localai-tests, one that resists SIGTERM and carries a large writable layer so the reap is slow:

cat > Dockerfile.race <<'DOCKER'
FROM busybox
CMD ["sh","-c","trap '' TERM; dd if=/dev/zero of=/big bs=1M count=800 2>/dev/null; while true; do sleep 1; done"]
DOCKER
docker build -q -t race-test -f Dockerfile.race .

# old teardown
docker run -d --rm --name r race-test
sleep 4
docker stop -t 1 $(docker ps -q --filter ancestor=race-test)
docker rmi race-test          # conflict, 5/5 trials

# new teardown
docker run -d --rm --name r race-test
sleep 4
CONTAINERS=$(docker ps -aq --filter ancestor=race-test); \
  [ -n "$CONTAINERS" ] && docker rm -f $CONTAINERS
docker rmi race-test          # succeeds, 5/5 trials

The empty-argument bug: docker stop $(docker ps -q --filter ancestor=nonexistent) exits 1; the replacement exits 0.

For the workflows, the E2E Backend Tests check on this PR should go green, and no tmate step should run because this PR carries no ci-debug label.

Not in scope

PR #11800 also has failing bonsai backend builds. Those are deterministic, not this flake: the fork bump makes json resolve to common_json, which has no parse_error, while backend/cpp/llama-cpp/grpc-server.cpp still catches const json::parse_error&. That needs its own change.