Transcript compaction rewritten against the v5 storage interface

A new performCompactTranscriptV5 compacts the session transcript via the storage interface, using paged forward reads and a replaceRecords call guarded by an ifUnchangedThrough precondition with atomic publish, rather than rewriting the transcript file directly. It bails out with specific reasons when conditions aren't safe: snapshot_mid_line for a torn tail, shared_inode for a hard-linked target, and source_changed when it loses a race. It also doubles its own backstop threshold when a compaction only reclaims a small amount of space, and reports bytes before and after on success.