A 'poison message' always crashes the consumer of a Kafka partition. Why is this worse than with SQS-style queues?