I suppose one could use a round-robin sharding approach like you mention, but it goes against Kafka's design, and it's not necessary.
Kafka divides a queue into partitions. Each partition is a completely independent silo. When you publish messages, Kafka distributes them across partititons. When you read messages, you always read from a partition.
This means partitions are also the unit of parallelism: You don't want multiple workers on a single partition (because of the labour division problem you mention). Rather, Kafka expects you to have one partition per worker.
This is more elegant than it sounds if you're coming from something like RabbitMQ. Partitions (ie., queues) in Kafka are append-only and strictly linear; unlike RabbitMQ, you can never "nack" a message in a way that results in the message ending up at the back of the queue and thus violating the original message order. Rather, Kafka expects each consumer to maintain its "read position" in the queue. Failure handling, then, is simply a matter of winding back the read position. And unlike RabbitMQ, there's less need for complicated routing, dead-letter exchanges and so on, because rather than move messages around, you're just moving a cursor.
Of course, message order is only preserved within a single partition; if you publish messages A, B and C and you have 3 partitions and 3 workers, then in a real world, messages may be processed in the order C, B, A. That sounds bad, but then other queue solutions such as Que or RabbitMQ suffer from the exact same problem: If you run 3 workers against one queue, your queue may supply each worker with messages in the right order, but there's no guarantee that they will be processed in that order. The only way to guarantee ordering is to have just one worker per queue, using some kind of locking (RabbitMQ does support "exclusive" consumers). But then you don't get any parallelism at all. So I think Kafka's solution is quite sane, even if it's more low-level and less developer-friendly than AMQP.