ci: Make the smoke test wait for health checks, and check more - #1487

Makes the smoke test wait for health checks instead of sleeping, and check more. One change reaches customers: three services get a longer health check start period.

Smoke test

• Runs docker compose up --wait --wait-timeout 300 instead of sleep 90. It returns once every container is running and healthy, and fails if one reports unhealthy.
• Pulls images in parallel first, with up to three attempts, because Docker Hub times out now and then.
• Counts the expected containers from the Compose file instead of a hard-coded 25.
• Fails if any container restarted.
• Saves the logs of every container on failure, including the ones that exited.
• Uses docker compose (v2) instead of docker-compose (v1).
• Adds these to the smoke test section of AGENTS.md.

What customers will see

searcher-0, precise-code-intel-worker, and syntactic-code-intel-worker get start_period: 300s on their health checks, the same as sourcegraph-frontend-internal. searcher-0 had none, and the two workers had 60 seconds.

None of the three is ready until sourcegraph-frontend-internal is, which waits for the migrator. Until then their health checks fail, so they could show unhealthy during a start, and --wait gives up on the first unhealthy container. With the start period they show starting until their first check passes.

Nothing else changes: Docker does not restart an unhealthy container, and no service waits on the health of these three.

Risk: low. One health check setting on three services; the rest is the test.

Checklist

• All images have a valid tag and SHA256 sum (no image changes)

Test plan

• Build 16954 passed, the smoke test job took 404 seconds
• Without the start period, --wait fails: build 16853 with container searcher-0 is unhealthy, and build 16905 with container precise-code-intel-worker is unhealthy
• shellcheck test/smoke-test.sh is clean

Closes FE-586