This is an attempt to support scaling from/to zero.
The basic idea is that we create a one-off "registration-only" runner pod on RunnerReplicaSet being scaled to zero, so that there is one "offline" runner, which enables GitHub Actions to queue jobs instead of discarding those.
GitHub Actions seems to immediately throw away the new job when there are no runners at all. Generally, having runners of any status, `busy`, `idle`, or `offline` would prevent GitHub actions from failing jobs. But retaining `busy` or `idle` runners means that we need to keep runner pods running, which conflicts with our desired to scale to/from zero, hence we retain `offline` runners.
In this change, I enhanced the runnerreplicaset controller to create a registration-only runner on very beginning of its reconciliation logic, only when a runnerreplicaset is scaled to zero. The runner controller creates the registration-only runner pod, waits for it to become "offline", and then removes the runner pod. The runner on GitHub stays `offline`, until the runner resource on K8s is deleted. As we remove the registration-only runner pod as soon as it registers, this doesn't block cluster-autoscaler.
Related to #447
dbd7b486d2