fix(backup): serialize and atomically publish snapshots (salvage #69411)

[kshitijk4poor]

Salvages the backup-safety commit of #69411 by @hao Wang — cherry-picked with authorship preserved.

Context — what this fixes, for whom

Anyone with automatic backups + manual snapshots: run_backup, create_quick_snapshot, and the automatic full-zip path could run CONCURRENTLY (gateway auto-backup firing while a user runs /snapshot), interleaving SQLite safe-copies and half-written snapshot directories. The commit adds a shared cross-process backup lock (_backup_operation_lock, lockfile + BackupInProgressError) and atomic publication: snapshots build in a staging dir and rename into place, so a crashed backup can never leave a half-snapshot that restore would trust.

Scope decisions (bundle split — the PR carries 4 distinct concerns)

Kept: the backup commit (f0fcaf5: 241 lines backup.py + 105-line stability test file), the concern with the clearest bug and cleanest boundary.
Adaptations:

• One conflict: main's snapshot path gained a failed-DBs abort message after the PR's base; kept main's message with the PR's staging_dir cleanup.
• Stripped the track_activity("backup", ...) telemetry wrappers — they import hermes_cli.perf_diagnostics, a module introduced by a DIFFERENT commit of the bundle (e304e97) that isn't salvaged here. The locking (the actual fix) is fully preserved; the telemetry can ride with a future perf_diagnostics PR.
  Dropped for separate consideration (documented in the close comment):
• Desktop startup-queue + prewarm removal (74c99e1, 0409d51): removal of the hover-intent prewarm feature is a maintainer taste call, and the desktop stores drifted on main
• Model discovery latency + perf_diagnostics module (e304e97): touches tui_gateway/server.py + ws.py which this campaign just modified (perf(cold-start): mitigate the ~14s GIL stall during backend init (salvage #60807) #77814); needs its own review against the new warm-list

Verification

• tests/hermes_cli/test_backup_stability.py: 6 passed (concurrent-backup serialization, atomic publish, staging cleanup)
• Full backup suite: 46 passed
• Mutation check: revert backup.py to main → error; restore → 6 passed
• ruff clean

Closes #69411 (backup core superseded by this salvage — original author credited via cherry-pick authorship; remaining concerns documented for re-proposal).