Track worker occupancy in stats

[neob91-close]

The utilization stat counts only time inside task code, so a worker saturated with many short tasks reads well below 100% — the dequeue/scan/lock overhead between tasks is invisible. This broke utilization-based autoscaling for us: workers scaled down during a large backlog.

Adds time_idle (blocking waits for new work, measured via measure_idle() around the two wait sites), and emits time_overhead and occupancy = 100 * (total - idle) / total — the "can this worker take more work" signal. Purely additive: utilization and time_busy keep their meaning.

Caveat for consumers that monkey-patch StatsThread (custom stats adapters): the worker now calls measure_idle()/report_idle_*, so adapters need those methods.