fix(audio): a world-only scenario lasts the sum of its scenes

[LeadcodeDev]

Closes #341.

The fault

resolved_scenario_duration read "last scene's offset + its duration" from
scene_start_offsets. For a World view that function deliberately gives every
scene the same offset — the view's own window start — because world scenes
share one window for embedded-video audio extraction. So the total collapsed to
0 + last_scene.duration.

That total sizes the synth buffer, so a score event with no to ("play until the
end") stopped at the wrong end. The video rendered at the right length and the
muxed track was padded, so ffprobe showed nothing wrong — only listening, or
measuring, did.

The fix

The end cursor was already computed for both view types. It simply was not
returned. timeline_offsets now yields view_ends alongside scene_starts from
one walk, so the two cannot drift, and scene_start_offsets remains a wrapper
with its existing signature and semantics.

World scenes keep their shared window start — that is what embedded-video
extraction wants. Only the total changes.

Measured, not reasoned

The reproduction from the issue: three world scenes of 3 s + 3 s + 2 s, a kick on
every beat, rendered with the binary before and after.

	before	after
t=0.5s	−8.0 dB	−8.0 dB
t=1.5s	−8.0 dB	−8.0 dB
t=2.5s	−91.0 dB	−8.0 dB
t=4.0s	−91.0 dB	−8.0 dB
t=6.0s	−91.0 dB	−8.0 dB
t=7.5s	−91.0 dB	−8.0 dB

−91 dB is silence. It began at 2.0 s, the last scene's duration, exactly as
reported.

Four tests

• a_world_only_scenario_lasts_the_sum_of_its_scenes — fails against the old
  computation with got 2 for a scenario that lasts 8 s.
• world_scenes_keep_their_shared_window_start — pins the semantics the fix
  must not disturb.
• a_world_view_followed_by_a_slide_view_still_totals_correctly — passes
  before the fix too, which is precisely why this survived: appending a slide
  view was the known workaround, so any scenario that had one looked fine.
• a_trailing_empty_view_does_not_shorten_the_total — an empty view must
  neither add to nor subtract from the total. The old code continued past it;
  the new one filters it out, and this pins that they agree.

cargo fmt --all --check, cargo clippy --workspace --all-targets -- -D warnings
and cargo test --workspace (1518) all clean.

Note on comments

Written comment-free, per the codebase-wide rule being applied in #345. The
reasoning that would have been a comment is in the commit message and above.