Copy relocated leaf pages as bytes during compaction

[venkat1701]

Part of #50. Stacked on #73: merge that first; this PR's base then moves to main once the #73 branch is deleted.

Compaction decoded every moved leaf into objects, copied them, and encoded them again. That was most of the roughly 2 GiB a pass allocated, and the GC pauses that came with it were what was left of the commit tail after #73.

Now a moved page from a tree without nested references is copied as raw bytes:

• read it
• verify it against its old page id, so a corrupt page fails instead of being re-sealed with a fresh, valid checksum
• check the height in its header; only leaves are copied
• re-stamp the header with the new page id and checksum, log it and write it, the same way the Materializer does

Branches and leaves that hold nested trees are still rebuilt. This also covers nested single-leaf trees (most small edges' membership trees), which are reached through a reference whose height the walker doesn't know; the page header tells it. Old and new pages are now paired after the copies are written, since copied pages never go through the node cache.

Results, HStore only, 3 alternating runs each, #73 against #73 plus this PR, load 9 to 21, all checksums identical:

during a full compaction	default segments	16 MiB segments
allocated	1,783 → 1,244 MiB (-30%)	2,141 → 1,459 MiB (-32%)
GC pauses	1,326 → 340 ms (-74%)	720 → 270 ms (-63%)
commit p99	692 → 267 ms (-61%)	348 → 195 ms (-44%)
commit max	771 → 343 ms	418 → 271 ms
compaction time	3.8 → 2.3 s	4.1 → 2.3 s

Space reclaimed and disk size after compaction are unchanged.

I also tried capping background passes at 64 MiB of live data. With 16 MiB segments that made passes bigger (16 to 17 segments instead of 1), and the mixed.write tail got worse. Real bounding needs moving part of a segment per pass, so I dropped that change.

Tested:

• full ./mvnw install passes (103 tests); each commit compiles on its own
• all 11 CompactionTest cases pass, including the snapshot-to-swap merge, two branches with reclaim and reopen, and the crash matrix of the swap
• new aCorruptPageIsNotCopiedWithAFreshChecksum: a membership leaf corrupted on disk (while its decoded copy is still cached) makes relocation fail. I checked that it fails if verification is skipped.

[venkat1701]

Closing this. #81 replaced tree-rewriting compaction with moving pages through the page directory, so most of this stack no longer applies. A review also found two data-loss bugs in #73 (a retirement generation older than a commit that still used the moved pages, and bulk-loaded pages left behind in a victim). The pieces that still help, the posting list decode, records reused by mapRefs, and the checkpoint crash test, are in #82.