[fix][broker] Move pending acks cleanup to selected mark-delete callbacks

[nodece]

Motivation

Pending-ack cleanup was previously executed in the readMoreEntries dispatch path. This introduced unnecessary overhead in a hot path and also risked concurrent access to the same pending-ack entries from both dispatch and ack threads.

Cleanup is only required when messages are removed through mark-delete advancing operations such as expiry, skip, or clear-backlog. Ack-triggered mark-delete operations are much more frequent and should avoid carrying additional cleanup work.

Modifications

• Introduced atomic remaining-unacked tracking for both non-batch and batch messages.

  • PendingAcksMap now stores and updates the actual remainingUnacked count instead of relying on batch size–based accounting.
  • Added atomic helper methods such as updateRemainingUnacked and removePendingAckAndGet to simplify and centralize unacked-message accounting.
  • Removed legacy logic that recalculated unacked counts from cursor state or batch size.

• Refactored pending-ack cleanup to leverage the existing markDeletePositionMoveForward hook.

  • Removed pending-ack cleanup from the readMoreEntries dispatch path.

  • Cleanup of pending acks and redelivery state is now triggered only after mark-delete advancement operations such as:

    • message expiry
    • skip entries
    • clear backlog

  • Ack-triggered paths do not invoke markDeletePositionMoveForward; they continue using afterAckMessages to avoid additional overhead on the high-frequency ack path.

• Updated pending-ack iteration and removal paths to consistently use the maintained remainingUnacked count directly, improving correctness and concurrency handling for partial batch acknowledgments.

[lhotari]

LGTM, follow up issue #25613