vello_cpu: Add fast path for pixel-aligned alpha-mask image fills

[nicoburns]

Summary

Cached glyphs are drawn as image fills from the glyph atlas with the following settings:

• Tint { mode: AlphaMask }
• Pad extend
• an identity-scale paint transform with an integer offset

Before this change, those fills took the generic painter path:

1. sample RGBA into paint_buf
2. run a separate apply_tint pass
3. composite the buffer

That path was ~18% of single-threaded cached rendering.

This PR adds a fast path to Fine's EncodedPaint::Image arm. It is taken when all of the following hold:

• the tint mode is AlphaMask
• both extends are Pad
• EncodedImageExt::integer_translation() returns Some((dx, dy)), i.e. the transform is exactly [1, 0, 0, 1] with near-integer offsets

On that path:

pad_alpha_mask(pixmap, x + dx, y + dy, width, strip_alphas, &mut mask_buf)
    // reads atlas alpha into strip-alpha layout (SIMD 4×4 transpose),
    // multiplied by the strip coverage; out-of-bounds uses Pad semantics
alpha_composite_solid(dest, tint_color, Some(mask_buf))   // or T::blend(...) for non-default blend / masks

The converted tint colour is cached across commands. Everything else still uses the existing painters.

A unit test compares pad_alpha_mask against a scalar reference, including offsets outside the image.

Part 3 of 3 of the CPU glyph-cache speedups. It conflicts with the Fine image-resolve cache PR in fine/mod.rs; whichever merges second needs a trivial rebase. On its own it speeds up single-threaded rendering. Multithreaded scaling additionally needs the resolve-cache PR, because this path still resolves the image, and so bumps the shared Arc refcount, on every command.

Measurements

Throwaway probe: 1600×1200 target, ~9.3k glyphs, 14px, atlas_cache(true), mean of 60 warm frames on a noisy VM. Render time, main → this branch:

threads	main	this branch
1	4.50 ms	2.73 ms
4	6.32 ms	5.74 ms (still contended, see above)

All three PRs combined, at 14px:

• 1 thread: 7.3 → 5.5 ms total (render 3.35 → 2.56 ms)
• 8 threads: 10.8 → 3.9 ms total (render 7.64 → 1.05 ms)

For comparison, uncached rendering at 8 threads is 8.0 ms total.

Probe output is pixel-identical to main, with the cache both on and off. cargo test -p vello_cpu -p vello_common passes.

Link to Devin session: https://dioxus.staging.devinenterprise.com/sessions/55344a6e161d48b4a3a8688fabe11d4c
Open in Devin Desktop: https://dioxus.staging.devinenterprise.com/desktop/session/55344a6e161d48b4a3a8688fabe11d4c?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