feat(frontend): expose cumulative prefix cache metrics through vllm

[xiaguan]

The frontend currently leaves vLLM prefix-cache queries and hits at zero even when a scheduler reuses cached tokens. Add cumulative request/query/hit counters to SchedulerMetrics and convert them to per-send deltas in both the watch and stepped bridges. Prefix-only steps send stats without waiting for output tokens; repeated snapshots contribute zero and coalesced snapshots retain every increment.

This provides the frontend contract for kern; existing model schedulers retain zero counters until they populate them.

Validation: 74 frontend unit tests pass, including all coalescings of four admissions, repeated snapshots, the watch publisher, and prefix-only stepped output. Frontend Clippy with all targets and warnings denied, plus workspace formatting, pass.