Metrics server: a stalled client blocks health endpoints; accept errors busy-loop

[lukashes]

Two availability defects in the observability HTTP server (http.zig).

• The serve loop is serial (one request per connection) and receiveHead has no read timeout. A client that connects and sends nothing blocks /healthz and /readyz for everyone. Under Kubernetes a stuck scraper turns into failed liveness probes and process restarts caused by observability itself. Add a read timeout or a per-connection deadline.
• Any accept error other than Canceled retries immediately and logs at debug only. A persistent error such as EMFILE spins at 100% CPU, invisible in production logs. Log at warn and back off between retries.