During initial cluster bootstrap the load balancer could get stuck in
`degraded` or `error` and never recover on its own. Two things combined
to cause this:

- The health monitor was created before any control-plane node existed.
  With a monitor but no healthy members, the LB reports `degraded`/`error`.
- The reconciler never requeued a non-running LB, so it never observed the
  LB recovering once the control-plane nodes came up. This was
  especially the case if it took a bit of time until control plane nodes
  were available.

Changes:

1. Requeue every 10s while the LB is pending (`degraded`/`error`)
2. Defer creating the health monitor until the control plane is
   initialized (at least one control-plane node up).
3. Move health monitor reconciliation to the end of `reconcileNormal`,
   after the load balancer and floating IP, since it depends on both
   being set.

Also fixes the `cloudscaleMachineToCluster` watch mapping, which filtered
on the wrong label (`MachineControlPlaneNameLabel` instead of
`MachineControlPlaneLabel`) and so never enqueued on control-plane
machine changes.