Poor handling of aggregated records in KinesisIO.read

The way the Kinesis source is implemented it doesn't play well with aggregated records.

Even using configuration options it's fairly hard to configure it in a way that becomes sufficiently performant.

One of the key issues is around bundle size & record queue size vs the number of aggregated records per message. These might, in certain situations, exceed the internal queue size by far unnecessarily blocking threads and requiring thread pools to be forcefully taken down.