perf(grid): semantics tree is rebuilt for every visible cell (~17 ms per result set, hidden from FrameTiming)

[ZhuchkaTriplesix]

Summary

The data grid creates semantics for every visible cell. On Linux the embedder keeps the semantics tree enabled at all times (SemanticsBinding.instance.semanticsEnabled == true in a normal session), and on Windows it is enabled whenever any UI Automation client is attached. Rebuilding that tree is one of the largest per-frame costs of the grid, and it is invisible in FrameTiming.buildDuration because flushSemantics runs after Render, while still blocking the UI thread before the next frame.

Evidence (profile, Linux, window 2056×2616)

benchmark/grid_perf_bench.dart, engine timeline totals:

Scenario	SEMANTICS	LAYOUT	BUILD
New result set every 400 ms (5000×20), current	565 ms / 31 frames	543 ms	288 ms
Same, rows wrapped in ExcludeSemantics (experiment)	38 ms / 33 frames	542 ms	281 ms

≈ 17 ms of semantics per new result set. Horizontal scroll: ≈ 1.9 ms of semantics per frame (555 ms / 298 frames).

Scope

• Give each grid row one semantics node (e.g. a label joining header + value pairs, or at least the row index and values), instead of one node per cell with gesture actions.
• Keep the grid usable with a screen reader (row-level reading; cell actions via the existing keyboard shortcuts).
• Re-measure with MODE=rerun and MODE=scroll, including the SEMANTICS timeline phase (not only FrameTiming).

Out of scope

Disabling semantics globally (accessibility must keep working).

[ZhuchkaTriplesix]

Resolved by #996 (already merged into dev; not auto-closed since this repo's default branch is main and PRs merge into dev).