tiered compaction: refactor and re-enable iterator based k-merge

[arpad-m]

PR #7661 has switched to a completely in-memory based k-merge due to assertion errors in the iterator based k-merge. This has a non-zero memory impact bounded on the number of delta layers considered at a time multiplied by the maximum layer size. Ideally, we'd find and fix the bugs in the k-merge implementation.

Also, tests for the k-merge would be very helpful.

@skyzh has expressed interest in this, or at least in refactoring the k-merge.

[problame]

Meeting notes: Chi authored #7760 which fixes this. Still missing test case.
Should be able to use the test that failed before we switched to buffered kmerge.

Time budget to understand with what was wrong with the prior kmerge: 1/2 day.
(NB the failure might now have been due to a bug in the prior kmerge impl, but the load/unload functionality)