perf(chat): virtualize the timeline by segment, not by turn

[Tryanks]

Why

Thread 9553a8bb… (7.6 MB, 19k events, 13 turns) is unusable on a phone. The transport work (compressed wire, 4 KiB output previews, 512 KiB pages) and the list virtualization both key on the turn, and this thread has one still-running turn holding 985 items (345 commands, 224 assistant notes, 179 reasoning items) that folds into 449 segments. One list row = one turn, so every frame rebuilt those 449 segments, and Markdown residency counted its 8-turn margins over a 13-turn thread, keeping essentially every document parsed.

Reference: Zed's agent thread view (crates/agent_ui/src/conversation_view/thread_view.rs) gives GPUI's ListState one row per thread entry, never per turn, and splices rows in place as entries land.

What changes

• One row per segment. TimelineRow (was TurnListItem) is a segment of a turn: a message, an error, a divider or a Work Log run. The turn's trailer (plan card, changed files, liveness/finished time, pending steers) renders on the turn's last row; a turn without segments keeps one empty row. segment_entries now also reports each segment's entry range so indexing and rendering fold the same way.
• Row sync is an ordered identity diff. A row is identified by its first entry, so a run keeps its measured height while it streams. Rows that appear are spliced in place (ListState::splice keeps the anchor), rows above everything are the existing Prepend path with its history reservation, and rows that vanish or reorder still reset, except inside the partial first turn. A steer accepted mid-turn is now an in-place insertion rather than a reset.
• Residency counts rows. Build/evict margins and the tail pin are in rows; a running turn is no longer pinned wholesale (it was pinning all 150+ documents of the running turn), only the tail rows are.
• No clone on refold. Timeline::fold_stored folds the client's held records by reference; each history page used to clone the whole held log (≈2 MB here) before refolding.

Turn-keyed behaviour is unchanged: expansion keys, open_diff_for_turn, rewind, palette jump and the highlighted turn all resolve through rows_of_turn.

Tests

• a_long_running_turn_virtualizes_into_one_row_per_segment (model): a 200-step running turn indexes to 402 rows with the live/last-assistant flags on the right rows, and the next command grows the last row in place.
• a_long_running_turn_parses_only_the_markdown_near_the_viewport (view): a single 150-note running turn opens with fewer than 60 documents resident and note-0 not parsed. Fails before this change (151 resident).
• List-sync tests now assert splices; residency tests use rows; chat layout tests updated for three rows per synthetic turn.

Checks run locally: cargo fmt --all --check, cargo clippy --workspace --all-targets --locked -- -D warnings, cargo nextest run --workspace --locked. Not exercised: iOS/Android/Web builds (CI), and a real phone against the affected thread.