Fix race conditions with binfmt and GPU Docker systemd plug-ins
Docker Machine provisions a runner with cloud-init, which
previously added systemd drop-ins for docker.service to
initialize binfmt for QEMU and install the GPU driver, if needed.
As seen in
gitlab-com/gl-infra/production#17835 (closed), we
can run into a race condition where docker.service gets enabled
before cloud-init gets a chance to write those files. As a result,
those drop-ins don't get run without calling systemctl reload daemon-service and systemctl restart docker.service.
We don't want to call systemctl restart docker.service in
cloud-init because that could abruptly terminate the Docker
connection, as we saw happen in
gitlab-org/ci-cd/shared-runners/infrastructure#203 (closed). Instead,
convert the systemd drop-ins to full-fledged services that are
manually started and run after docker.service.