ci: Run the smoke test VM with gcloud instead of Vagrant - #1488

Replaces Vagrant with a gcloud script for the smoke test VM. .buildkite/smoke-test-vm.sh creates the VM, copies the repository to it over SSH, runs the same tests, and deletes it.

• No Vagrant. Removes .buildkite/vagrant-run.sh, test/Vagrantfile, and test/servers.yaml.
• Faster. No Vagrant plugins to install on every run, and the job does not wait for the VM to be deleted. The smoke test job took 333 seconds here, and 451 seconds with the same tests on Vagrant (ci: Check every Compose file and shell script, and test Jaeger and the executor #1489), one run each.
• 1 hour limit. GCP deletes each VM an hour after it starts (--max-run-duration=1h), so a job which dies cannot leave a VM running. The step times out after 30 minutes.
• Cleans up leftovers. Each job first deletes smoke test VMs and unattached disks older than an hour.
• 20 GB boot disk, set explicitly. Before, the VM took the image's size.
• Secrets. Each job fetches the builder key into its own temporary directory and removes it; before, every job on an agent shared /tmp/e2e-builder.json. The Docker Hub password goes to the VM over stdin, and is no longer written to /root/.profile.
• Retries the SSH connection, on the same VM. In build 16777 the VM stopped answering while the repository was being copied to it. The script retries that copy every 15 seconds, up to 10 times, when ssh exits with 255. It never creates a second VM, and never retries a test failure.
• Updates AGENTS.md for the new script.

Risk: low. CI only. The VM image, machine type, network, and service account are unchanged.

Things to know

• The script calls gcloud beta, because the agents have gcloud 417, where --max-run-duration is only in the beta track.
• The agent queue is still named vagrant. Renaming it is not part of this PR.
• A connection lost while the tests are running still fails the job, because the script cannot tell how far the tests got.
• This should be backported to 8.2.x, so that Vagrant can come off the CI agents. It does not apply there alone: it needs ci: Skip the smoke test on branches which only change files it does not read #1484, ci: Make the smoke test wait for health checks, and check more #1487, and ci: Check every Compose file and shell script, and test Jaeger and the executor #1489 first, or a version made by hand.

Checklist

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

Test plan

• Build 16957 passed: both tests ran on the gcloud VM
• With stub gcloud, ssh, and curl commands, each of these creates one VM, requests one deletion, and removes the key files:
  • a passing run
  • a failing test: exits 1, no retry, logs copied back
  • a copy which fails twice with 255, then works: two retries, then the tests run
  • a copy which never works: ten retries, exits 255, the tests do not run
• The retry has not run in Buildkite: I could not make a real VM drop off the network on demand

Closes FE-584