macOS: pace the window's frame to the display, and never wait for a drawable on the main thread

[albanmgicquel]

On macOS the window host polled (ControlFlow::Poll, the frame in about_to_wait). Under FIFO presentation the frame spent most of its time in get_current_texture, which is nextDrawable waiting for vsync, on the main thread: 90% of the run in a game we measured.

Window managers with a click-filtering event tap, such as Magnet or BetterTouchTool's snap areas, hold each mouse-down while they ask the clicked app about its window over Accessibility. That request is served by the main run loop, one per turn, and each turn was a blocked frame. The result: every click, and the start of every title-bar drag, waited about 0.5 s at 60 Hz (less at 120 Hz), while the host's own timings looked perfect. This is rust-windowing/winit#1737; Godot ships a workaround that detects Magnet by its bundle id.

This change, on macOS:

• The event loop sleeps in WaitUntil until the next frame is due, one monitor refresh after the last. The frame is requested with request_redraw and runs in RedrawRequested.
• A worker thread acquires each drawable as soon as the previous frame is presented. A due frame whose drawable isn't ready yet is put off and checked again every 1 ms, at most 100 ms, then runs and waits as before.
• Between frames the run loop sleeps, so Accessibility requests are answered right away.
• It uses only wgpu's public API (Surface and SurfaceTexture are Send + Sync). A reconfigure first settles the acquisition in flight.

Measured with an equivalent diagnostic build, with Magnet running:

	Before	After
Main thread blocked in acquire	90%	0%
Mouse-down age on arrival	~500 ms	max 16 ms (the tap's own cost)

Alternatives tried and rejected:

• Timer pacing without the gate: still blocked up to a frame and held clicks up to ~130 ms.
• A CADisplayLink: no better than the timer once gated (and blocking without the gate).
• One frame of latency: halves the frame rate.
• Counting presented drawables: needs a wgpu-hal patch.

Other platforms and the benchmark's interactive view keep the polled loop. One caveat, documented: wgpu-hal's occlusionState read now runs on the acquisition worker thread.

Tests: pacing.rs unit tests (deadline, gate, its bound, no catch-up after a late frame, parked retry, the acquisition worker asking rather than waiting, settling, shutdown). cargo fmt, cargo clippy -p tecs-winit-host --all-targets -D warnings, cargo test -p tecs-winit-host (110) and nupp test pass. Verified by hand on macOS with Magnet running. Documented in a README design record and docs/gfx/benchmarks.md.

Not run here: nupp task docs-check (this toolchain lacks the lunamark rock).