Stop SSE reconnect resets from panicking; add staleness metrics

[ewokndor]

Symptom:

Recurring error logs {"StreamID":N,"Code":2} ~every 20 min, with the service panicking (and restarting) roughly hourly.

Root cause:

hermes.pyth.network sits behind Cloudflare, which periodically resets the HTTP/2 stream (INTERNAL_ERROR, connection left intact) every ~5–12 min by design — confirmed by reproducing it with plain curl (cf-ray header =
Cloudflare). Two client bugs turned that benign reset into noise + crashes:

1. r3labs' default reconnect backoff has a 15-min MaxElapsedTime measured from subscription start, so after the stream lived >15 min the next reset surfaced as a fatal error instead of reconnecting.
2. Our retry counter never reset across independent disconnects, so every 3rd reset hit maxRetries and panicked, crash-looping the pod ~hourly.

So every log line was a real disconnect (~20 min apart); the num_retries 1/2/3 ladder was independent disconnects accumulating toward the panic, not retries of one event.

What we changed (hermes/):

• Reconnect through Cloudflare resets indefinitely (ReconnectStrategy with MaxElapsedTime=0) — recoverable, logged at info.
• Removed the panic/retry-counter; unrecoverable errors now log at error and re-subscribe with capped backoff (no more crash-loop).
• Added health signals for metrics-based alerting:
  • LastStreamUpdate() — global stream liveness (transport health).
  • LastFeedPublishTime(feedID) — per-feed staleness from Pyth's publish_time, catches a single feed freezing even while the stream is alive.

Alerting:

gauges live in the caller. Export time.Since(LastFeedPublishTime(id)) per feed as NoPriceUpdateSince{feed=...}; one Prometheus rule (> 30s) fans out per feed automatically. Plus a global backstop from LastStreamUpdate().

Net effect:

error-log spam and hourly panics gone; reconnects are visible but quiet; stale prices now trip a metric alert regardless of why updates stopped.