fix(prometheus): serve scrapes from cached background poll (#1013)

Decouples Prometheus scrape cadence from upstream UniFi API calls so a
429 backoff loop on the controller side no longer stalls /metrics. The
output plugin now owns a 60s background poller (configurable) whose
result is served from an in-memory cache. Concurrent /scrape requests
for the same target are coalesced via singleflight to prevent a noisy
scraper from multiplying upstream load.

Adds two new metrics so operators can detect cache staleness and
refresh failures independently:
- unpoller_prometheus_cache_age_seconds
- unpoller_prometheus_refresh_failures_total

Background goroutine recovers from panics so a malformed input payload
no longer silently kills refreshes.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

examples/up.conf.example

@@ -35,6 +35,10 @@
   report_errors = false
   ## Record data for disabled or down (unlinked) switch ports.
   dead_ports = false
+  # How often the background poller refreshes the cache served to /metrics.
+  # Decouples scrape cadence from UniFi API calls so 429 backoff loops no
+  # longer block /metrics. Default: 60s. Values below 15s are clamped to 15s.
+  interval = "60s"
 
 [influxdb]
   disable = false