[sparse strips] Support elliptical corner radii in blurred rounded rectangles

[nicoburns]

Summary

Stacked on #1. Extends the analytic blurred rounded rect paint from scalar per-corner radii to per-corner elliptical radii (CSS border-radius with / syntax), across vello_common, vello_cpu, and vello_hybrid.

New public type in vello_common::blurred_rounded_rect (kurbo's RoundedRectRadii can't express elliptical corners):

pub struct CornerRadii {
    pub top_left: Vec2,     // x radius, y radius
    pub top_right: Vec2,
    pub bottom_left: Vec2,
    pub bottom_right: Vec2,
}
// From<f64> and From<RoundedRectRadii> preserve the old call sites via `.into()`

fill_blurred_rounded_rect on vello_cpu::RenderContext and vello_hybrid::Scene now takes CornerRadii (breaking, follows on from #1's change).

Field math

The isotropic superellipse corner term is replaced by an anisotropic one. Per corner we encode blur-adjusted axis radii r1_x = hypot(rx, 2σ), r1_y = hypot(ry, 2σ), an exponent derived from the corner's geometric mean radius (which reduces to #1's exponent for circular corners), plus precomputed weights w_x = r1_x^-p, w_y = r1_y^-p. In the fine shader (same code in vello_cpu SIMD and render.wesl):

a = x1^p; b = y1^p
d_pos = (a + b)^(1/p)                                  // unchanged shape of the field
r_eff = ((a + b + ε) / (a·w_x + b·w_y + ε·mean_w))^(1/p) // direction-dependent effective radius
d     = d_pos + d_neg - r_eff

r_eff is the radius of the anisotropic superellipse in the direction of the sample point — it interpolates between r1_x (on the horizontal midline) and r1_y (vertical), so the Gaussian falloff tracks the local corner curvature instead of assuming a circle. When r1_x == r1_y the shader selects r1_x directly, so circular corners take exactly the same value path as before — all existing uniform and non-uniform snapshots pass byte-identical, on both CPU and hybrid.

GpuBlurredRoundedRect grows from 6 to 9 texels (r1_x, r1_y, exponent, w_x, w_y corner vectors).

Caveats

This remains an approximation: the anisotropic field's iso-contours are not uniformly spaced Euclidean offsets, so blur falloff around highly eccentric corners is heuristic rather than exact. It is well-behaved for moderate eccentricity and improves as σ grows; the fully-elliptical pill snapshot (40×25 radii) renders correctly.

Testing

New snapshot tests (CPU + hybrid/wgpu): blurred_rounded_rect_elliptical_radii, ..._elliptical_radii_std_dev, ..._elliptical_pill, inverse_blurred_rounded_rect_elliptical_radii. An elliptical example was added to the blurred_rounded_rect example scene. cargo fmt / clippy --workspace --all-targets clean; full blurred_rounded_rect filter: 192 passed, 0 failed.

Link to Devin session: https://dioxus.staging.devinenterprise.com/sessions/f0e29f278fc544ebb7a91bfeed0bcc85
Requested by: @nicoburns

[staging-devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR that start with 'DevinAI' or '@devin'.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring