changefeedccl: leverage kafka data loss detection to perform some action, such as replaying the changefeed from the last known good point

franz-go, the kafka library powering our v2 kafka sink, has the ability to detect unavoidable data loss due to kafka-side operational issues. when this happens, we currently (in the v2 sink) warn. we should instead try to mitigate the data loss by replaying messages.

Jira issue: CRDB-40338