feat: drain run backlog back-to-back with a job capacity guard

[tfelix]

The run-controller previously created at most one job per polling interval, so a backlog of runs drained at 1 run / interval. It also had no concurrency limit: it would claim a run, fail job creation when the namespace was full, and report the run as FAILED.

Process queued runs back-to-back each polling cycle until the backlog is empty, and gate fetching on a best-effort capacity check so we only claim runs we are fairly sure we can place:

• Add maxConcurrentJobs config (default 20; negative = unlimited).
• Count unfinished runner jobs (CountActiveJobs) and drain only up to the remaining budget; skip the cycle when at capacity or when the count cannot be determined.
• processNextRun now returns a result so the drain loop knows when to stop (no run, capacity reached, or failure).
• Add run_controller_jobs_at_capacity_skips_total metric.
• Introduce JobManager interface to unit-test the control flow; add tests for draining, capacity limits, and availableCapacity branches.