xkafka: Producer.AsyncPublish causes goroutine leak when Start/Run is not called

[sonnes]

Summary

The Producer struct requires Start() or Run() to be called to drain the internal events channel (delivery reports). However, AsyncPublish works without Start() being called — it silently pushes messages via ProduceChannel() without any compile-time or runtime indication that the event loop is missing.

Problem

When Start() is not called:

1. The events channel (producer.go#L17) fills up with unhandled delivery reports
2. The confluent-kafka-go internal poller blocks because it cannot push to the full events channel
3. ProduceChannel() backs up because the poller is blocked
4. Any goroutine calling AsyncPublish blocks forever on p.kafka.ProduceChannel() <- km
5. This results in an unbounded goroutine and memory leak

Reproduction

Call NewProducer() followed by AsyncPublish() repeatedly without ever calling Start() or Run(). Observe goroutine count growing without bound.

Suggested Improvements

• Document the requirement: The current doc on Start() only says "starts kafka event handling" without explaining it is mandatory for AsyncPublish to function correctly. Add explicit documentation that Start()/Run() must be called when using AsyncPublish.
• Return an error: Consider returning an error from AsyncPublish if Start() has not been called yet.
• Auto-start: Alternatively, auto-start the event loop lazily on the first AsyncPublish call via sync.Once.