internal/compact: DeletableValueMerger needDelete path corrupts iterator state

## Summary

When a merger implementing `base.DeletableValueMerger` returns `delete=true` from `DeletableFinish` during a compaction, the compaction iterator's `needDelete` handling elides the merge result via `continue` without performing the `skip/position` bookkeeping that the normal returned-key path performs. Depending on arrangement this causes either:

- Silent data corruption: the just-consumed base SET (and any older same-key entries the pending skip was supposed to elide) are re-processed as fresh keys, resurrecting a value for a key the merger declared non-existent; or
- A panic: compaction iterator has skip=true, but iterator is at iterPosNext on a subsequent Iter.Next() call.

Reproduced on `v2.1.4`, `v2.1.5`, `v2.1.6`; the relevant code is unchanged on master.

Jira issue: PEBBLE-1469

Start in internal/compact at the compaction iterator's needDelete handling and trace the DeletableFinish path alongside the normal returned-key path. Reproduce the report on v2.1.4–v2.1.6 or master, then verify that skip/position bookkeeping remains consistent, with no resurrected values or iterPosNext panic.

Written by the indexing model from the issue text.