Emit client-observed benchmark runner metrics

[robholland]

Summary

• Emit client-observed benchmark_runner_invocations_{started,completed,failed}_total and benchmark_runner_invocation_duration_seconds{outcome} from the Temporal load runner.
• Register them on the existing Prometheus endpoint alongside SDK metrics; no new scrape target is needed.
• Buckets are dense from 5 ms to 60 s and extend to 10 minutes for slower workloads.
• Count completion and successful latency only when the runner waited and observed workflow completion; a successful -w=false submission is only a start.