Review Kafka exactly-once checkpoint data-loss fix

The Kafka exactly-once sink could lose the first record when checkpoint preparation captured transaction state before asynchronous send registration completed. The PR also needed precise documentation of how checkpoint failure and replay behave across engines.

Reviewed the latest Kafka transaction and checkpoint lifecycle, confirmed the flush-before-state-capture and async-send failure handling addressed the race, and suggested engine-neutral documentation language for checkpoint failure and replay.

• Strengthens review confidence for a high-impact data-loss fix in SeaTunnel's Kafka exactly-once sink path.
• Clarifies that failure handling prevents an unsafe transaction commit and relies on replay from the last completed checkpoint, without overstating abort behavior for every engine.
• Documents a review contribution separately from authored-code PRs so the portfolio does not overclaim authorship.

• Review comment recorded no code correctness blocker after checking the transaction and checkpoint lifecycle.
• Follow-up discussion confirmed the documentation precision issue and framed it as non-blocking.
• Merged after maintainer approval; the landed commit is not authored or co-authored by Goutam, so this card is labeled as review contribution.