feat(metrics): serve prometheus metrics at /metrics

[jamiehdev]

shadowstep now serves Prometheus metrics at GET /metrics in text format 0.0.4, so operators can scrape request, cache and origin counts. /health reads the same counters, so the two cannot drift, and its JSON shape and values stay the same.

Changes

• The prometheus crate with default features off produces the exposition. It writes format 0.0.4 directly and adds one package to the lock file. prometheus-client writes only OpenMetrics text, whose counter TYPE lines omit _total and so mismatch the 0.0.4 media type.
• The metrics are shadowstep_requests_total{route,cache}, shadowstep_responses_total{route,status_class}, shadowstep_origin_requests_total{kind,outcome}, shadowstep_origin_body_idle_timeouts_total{kind}, the shadowstep_origin_response_seconds histogram (5 ms to 30 s), two cache counters and two cache gauges. No label holds a path.
• A body that goes idle counts in its own counter, because it follows a response that origin_requests_total already counted by status class.
• --metrics-addr (METRICS_ADDR) moves /metrics to a separate listener that serves nothing else. The proxy listeners then forward /metrics to the origin. Without the flag, /metrics stays on the proxy listeners, as /health does. The README describes the exposure and the path shadowing.
• The Deployment gets no prometheus.io/scrape annotations, because they are a scrape-config convention and not a standard.

Verification

• On main, all 11 new tests in tests/integration/metrics.rs fail, because the origin answers /metrics.
• One deliberate break for each metric family makes at least one test fail. The breaks cover wrong cache labels, status classes, origin outcomes and kind, a missing histogram observation, swapped gauges, and uncounted revalidations, refreshes and idle timeouts. They also cover a slash in a label, a counted scrape, and /metrics kept on the proxy listener.

Not run: a scrape by a real Prometheus server.