Granular runner dispatch telemetry by fleet

Dispatch telemetry for runner work now preserves the granular reason a dispatch cycle produced no work, such as an empty queue, lost claim race, or pod already in use, instead of collapsing all of those cases into a single outcome. The count and latency metrics are also tagged by fleet, so it is easier to distinguish a genuine per-shape dispatch stall from an idle warm pool simply polling for work.