feat(thumbnails): background thumbnail refresh + manual Update Thumbnails control

[thetalkingdrum]

Summary

Applying settings, a preset, a scan-setup change, or Batch Analysis to more than the
current frame writes straight to the DB for every other touched frame — but nothing
refreshed those frames' filmstrip thumbnails short of opening them by hand.

• A background pass re-renders and persists each touched frame's thumbnail without
  opening it.
• It never blocks Export, Stitch, HDR, or anything else — no progress dialog pops, no
  batch is refused because of it.
• Auto Crop All and Batch Analysis do share its render thread, so they pre-empt it
  immediately (within about one frame) instead of waiting it out — the frames that got
  cut short are automatically retried once they're done, with a status message so it's
  not silent.
• Diptych/half-frame rows are intentionally left out of scope for now: a whole-frame
  render can't reproduce how their canvas joins two half configs.
• A manual Update Thumbnails toolbar button (whole roll) and context-menu entry
  (current selection) let you force this on demand, for staleness an automatic trigger
  missed or predates one. The same control turns into Cancel while it's running —
  a very large accidentally-opened folder needs a way out, and unlike a real batch's
  pre-emption, cancelling discards the backlog outright instead of resuming it.

How it works

• desktop/workers/render.py: new ThumbnailRenderWorker, its own CPU-only
  ImageProcessor and decode cache, on the existing norm_thread.
• desktop/session.py / desktop/controller.py: a frames_edited_offscreen signal
  fired by every bulk-write path (Apply Settings, Apply Preset, scan setup, Batch
  Analysis, Apply Normalization Roll), dispatching refresh_thumbnails_for.
• _begin_batch pre-empts a running refresh only for the two batch types that actually
  share its thread (Auto Crop All, Batch Analysis) — everything else runs alongside it
  untouched.
• desktop/view/sidebar/files.py: the manual toolbar button and context-menu entry,
  routed through new request_thumbnail_refresh/cancel_thumbnail_refresh controller
  methods — same refresh_thumbnails_for plumbing underneath, no new backend path.

Heads up on overlapping open PRs

• fix(session): let a roll-wide reset undo an unwanted setting (#1047) #1118 ("Reset Roll to Defaults", now ready for review too) is a real, literal
  conflict: it inserts reset_roll_settings in session.py right after
  apply_preset_fields, the exact function this PR adds a
  frames_edited_offscreen.emit(...) block to the tail of — same hunk. It also edits the
  identical docs/USER_GUIDE.md right-click-menu sentence this PR does. Both are trivial
  to resolve by keeping both additions; whichever of us merges second does it by hand.
  Separately: this PR's frames_edited_offscreen signal is wired into Apply Settings,
  Apply Preset, scan setup, Batch Analysis, and Apply Normalization Roll, but not into
  fix(session): let a roll-wide reset undo an unwanted setting (#1047) #1118's reset_roll_settings — so a roll-wide or selection reset still leaves non-active
  frames' thumbnails stale after both land, unless reset_roll_settings is later wired to
  emit it too (or Film strip: two-column negative stats, stale-thumbnail indicator #1115's stale-dot approach lands and catches it structurally instead).

Test plan

• make lint / make type clean
• New unit tests for the worker, the controller wiring, every bulk-write trigger's
  signal emit, the manual control's dispatch/cancel behavior, and the real widget
  wiring (toolbar button, context menu)
• Full suite green (one pre-existing, unrelated Norwegian-keyboard tooltip failure,
  also fails on main)
• Manually verified in the app: Apply Settings to roll, Apply Preset to roll,
  exporting a frame mid-refresh, Batch Analysis pre-empting and resuming, opening a
  different non-active frame mid-refresh, a forced-Slide-mode frame's color
  surviving the background render unchanged, the manual Update Thumbnails button
  and context-menu entry, and cancelling a roll-wide run mid-flight

[thetalkingdrum]

Heads up: merge #1118 before this one. It touches the same hunk in session.py (reset_roll_settings lands right where this PR's frames_edited_offscreen.emit(...) goes, at the tail of apply_preset_fields) and the same docs/USER_GUIDE.md right-click-menu sentence.

When this rebases past #1118, fold reset_roll_settings into the wiring list here too. #1118 flagged it as a known limitation — a roll/selection reset leaves a stale thumbnail on non-active frames, the same gap Apply Settings/Apply Preset had before this PR — but the reset path itself isn't wired to emit the signal yet.

[thetalkingdrum]

Thanks for the note on the #1107 thread about coordinating with #1121's memory work — added in 51af764.

GPU needed no coordination: ThumbnailRenderWorker already uses its own ImageProcessor(use_gpu=False), so it never touches the GPU texture pool #1121 bounds.

RAM was the real gap: this refresh gets its own PreviewManager (following the same pattern the existing Auto Crop All worker already uses), so worst-case combined preview-cache footprint goes from 2x to 3x the fixed per-instance cap — and unlike #1121's prefetch path, nothing here checked memory pressure before starting a generation.

Fix: refresh_thumbnails_for now checks available_system_memory_bytes() against the same MIN_RAM_RESERVE_BYTES reserve #1121's prefetch admission uses (promoted out of prefetch_policy's private constant so both share one number), and defers + retries after 5s instead of starting a refresh under pressure. Falls back to proceeding if the OS can't report available memory, so it doesn't silently stop working on platforms without that introspection.