Mark annotated frames on the timeline

[emlcpfx]

Annotated frames now show as ticks on the timeline, so a reviewer scrubbing a shot can see where the notes are instead of hunting for them.

Ported from reviewapp's Timeline.set_markers, keeping its colours and shape:

• blue (#4fc3f7) — the frame holds comments
• purple (#ab47bc) — the frame holds only drawings

A frame with both reads as a comment: the words are the reviewable part, and the drawing marker mustn't hide them. (reviewapp arrives at the same result — its has_c and has_d branch returns the same blue as has_c.)

Markers are drawn under the playhead, so a note never hides the frame the reviewer is actually sitting on.

The bit that would have been wrong

Annotations are keyed by the player's LOCAL frame, but the timeline is a global range during playlist playback. Markers are mapped across with _timeline_frame_for_local, exactly as the [ / ] note navigation already does. Without it, the notes for one shot land on top of a different shot — and it would look plausible right up until someone tried to click one.

Keeping the markers honest

Comment add / delete / resolve and the sidecar load already funnel through single call sites, so the refresh hangs off those.

Strokes had no such hook: drawing changes what the timeline should show, and only the viewer knows it happened. So ViewerWidget gains an annotations_changed signal, emitted when a stroke is completed, undone, redone or cleared. Without it the markers would be right until you drew something, then quietly drift until the next source reload — the worst kind of wrong, because it looks fine.

Tests

10 new tests, 248 on the branch, compile clean.

The timeline is a plain QWidget (not the GL viewer), so the tests render it offscreen and probe the real pixels rather than trusting what the widget believes it stored:

• the two marker colours land on the right frames
• a both-kinds frame reads as a comment and isn't drawn twice
• an unannotated frame stays clean
• out-of-range markers (a stale frame from a longer clip) don't smear onto the track ends
• the playhead wins the pixel it shares with a marker

Also driven end to end through a real MainWindow with a decoded clip: adding a comment adds its marker, completing a stroke adds its marker, undo takes it away again, and deleting the comment clears it.

Branches off current main.

[D-Mad]

Superseded by #33. The replacement preserves the contributed timeline marker UI and adds playlist-wide aggregation, duplicate-shot mapping, stale-marker clearing, validated lightweight sidecar reads, and regression coverage. Thank you for the contribution.