vello_cpu/vello_hybrid: use rect fast paths for rect fills and clips

[nicoburns]

Summary

vello_cpu and vello_gpu 0.3 have rectangle fast paths (fill_rect, push_clip_rect; see linebender/vello#1916, #1924, #1926). AnyRender never used them, because every shape went through shape.into_path(..) and then fill_path / push_clip_path / push_clip_layer.

All three sparse-strip painters (vello_cpu, vello_hybrid, and vello_hybrid's WebGL painter) now check Shape::as_rect():

// fill
if let Some(rect) = shape.as_rect() { ctx.fill_rect(&rect) } else { ctx.fill_path(&shape.into_path(..)) }
// push_clip_layer (all three) and the clip of push_layer (hybrid / WebGL only)
if let Some(rect) = clip.as_rect() { ctx.push_clip_rect(&rect) } else { /* path */ }

vello_cpu push_clip_layer no longer always pushes an isolated clip layer:

if let Some(rect) = clip.as_rect()       { push_clip_rect(&rect) }   // pop_clip
else if ctx.is_multi_threaded()          { push_clip_layer(&path) }  // pop_layer
else                                     { push_clip_path(&path) }   // pop_clip

• Single-threaded: non-isolated clips match vello_hybrid and Skia, and are faster.
• Multi-threaded: isolated clip layers are kept for non-rect clips. The multi-threaded dispatcher flushes its pending batch on every push_clip_path / pop_clip, copies the clip into each batch, and rasterizes the clip on the main thread. In multi-threaded mode that costs more than the layer it saves. The check is RenderContext::is_multi_threaded() rather than the multithreading feature, because vello_cpu falls back to single-threaded when num_threads is 0.
• Pop bookkeeping: VelloCpuScenePainter gains a small layer_stack to choose between pop_layer and pop_clip.
• vello_cpu push_layer still passes its clip to push_layer as a path, because opacity and blend need a layer anyway.
• Strokes are unchanged, because stroke_rect has no fast path.

Vello itself falls back to paths for skewed or rotated transforms. A new vello_cpu unit test, rect_clips_and_fills, covers a rect clip nested inside a rect-clipped layer.

Benchmark

paint_bench cpu backend through Blitz, 1366×768@2x, 8-core VM. Each figure is the median of 4 runs of 60 frames, on a synthetic page of 600 rounded overflow: hidden boxes (non-rect clips):

	paint_scene	rasterize
single-threaded, layer → push_clip_path	36.3 → 33.2 ms	7.9 → 6.8 ms
multi-threaded, layer → push_clip_path	18.4 → 23.4 ms	2.6 → 2.2 ms

Multi-threaded therefore keeps the layer. Real pages (bbc, google, github, servo) showed no clear change, because their clips are almost all rects.

Link to Devin session: https://dioxus.staging.devinenterprise.com/sessions/45dbc6c083354777aaa43442ce8ef8ab
Open in Devin Desktop: https://dioxus.staging.devinenterprise.com/desktop/session/45dbc6c083354777aaa43442ce8ef8ab?variant=devin-insiders
Requested by: @nicoburns

[staging-devin-ai-integration]

I'll fix CI failures and address comments from users with write access that start with 'Devin'.

• Disable automatic comment, CI, and merge conflict monitoring