Make EphemeralRunnerReconciler create runner pods earlier

[mumoshu]

Tweaks the reconciler so that there are fewer requeues in the happy path, improving
runner startup time when dozens or hundreds of jobs are queued.

This is a pull request for the first commit mentioned in my comment at #3276 (comment).
It should fix, or at least alleviate #3276

How the improvement works

Let's simplify the ephemeral-runner reconciliation logic into generally 2 phases: (1) get token and (2) create runner pod.

This change improves the runner pod creation latency by making (2) happen earlier.

How it should work ideally

With three concurrent job startups, it should ideally be like this:

reconciliation #1 at t1: job 1 (1)
r #2 at t2: job 1 (2): 1 pod created until t2
r #3 at t3: job 2 (1)
r #4 at t4: job 2 (2): 2 pods created until t4
r #5 at t5: job 3 (1):
r #6 at t6: job 3 (2): 3 pods created until t6

How it works today

With the unnecessary requeues in the happy path, it had to be like this:

reconciliation #1 at t1: job 1 (1)
r #2 at t2: job 2 (1)
r #3 at t3: job 3 (1)
r #4 at t4: job 1 (2): 1 pod until t4
r #5 at t5: job 2 (2): 2 pods until t5
r #5 at t6: job 3 (2): 3 pod s until t6

You can see that the first two jobs take longer to start than the ideal scenario.
The more requeues in the happy path and concurrent N jobs you have, the longer the first N-1 jobs take to start.