vello_hybrid: Extend calculation for images is wrong

The following test case causes a huge mismatch between Vello CPU and Vello Hybrid:

#[vello_test(width = 100, height = 100)]
fn image_bilinear_probe_centered_pad(ctx: &mut impl Renderer) {
    let image_source = ctx.get_image_source(bilinear_probe_gradient());
    let image = Image {
        image: image_source,
        sampler: ImageSampler {
            x_extend: Extend::Pad,
            y_extend: Extend::Pad,
            quality: ImageQuality::Medium,
            alpha: 1.0,
        },
    };

    ctx.set_paint_transform(Affine::translate((47.5, 47.5)));
    ctx.set_paint(image);
    ctx.fill_rect(&Rect::new(10.0, 10.0, 90.0, 90.0));
}

Left is Vello CPU, right is Vello Hybrid:

Vello CPU correctly renders the right part in fully red (255,0,0), while Vello Hybrid renders it as (223, 0 32). The problem is that in Vello hybrid, we first apply the extend to the coordinates and then apply the -0.5/+0.5 offsets for bilinear sampling. However, how it should be done (and how it is currently done in Vello CPU) is that each position for bilinear sampling is extended individually. Unfortunately, fixing this would probably make image rendering even slower than it already is...

Start by locating and running the image_bilinear_probe_centered_pad test case. Compare the Vello CPU and Vello Hybrid image-sampling paths, focusing on Extend::Pad and the bilinear sampling offsets. Done means Vello Hybrid matches Vello CPU at the padded edge, including the fully red pixels shown.