Snap solid filled rectangles to whole pixels in the web (Canvas 2D) backend #638

Follow-up to #637.

Problem

#637 makes the Skia backend fill a solid, axis-aligned rectangle snapped to the nearest pixel edges without anti-aliasing, so thin barcode bars (e.g. TCPDF's x y w h re f bars, ~1.13 pt wide with ~1.13 pt gaps) stay solid and separate.

The web backend (pdf-graphics-web) does not do this. WebCanvasBackend::fill_path (crates/pdf-graphics-web/src/web_canvas_backend.rs) sends every path to Canvas 2D at its exact fractional coordinates, and Canvas 2D always anti-aliases fills. Thin bars therefore render grey-edged and narrower than in Skia/PDFium — most visibly at 1x zoom on standard-density displays.

Proposed change

Reuse the backend-agnostic helpers added in #637:

1. In fill_path, when there is no shader, compute the mapping from logical device space to backing pixels (it includes viewport.device_to_backing(), i.e. the devicePixelRatio scale).
2. Call CanvasPath::axis_aligned_rect(&to_pixels); if it returns a rectangle, take Rect::snapped_to_pixels().
3. Draw it with fillRect under an identity transform (restoring the transform afterwards). With integer edges, Canvas 2D's anti-aliasing produces no partial coverage.
4. Otherwise fall through to the existing path fill.

Snapping must happen in backing pixels, not logical device units; snapping before the DPR scale would misplace edges on high-density displays and could merge bars again.

Verification

• cargo xtask web-canvas --profile debug --serve --port 8080, open a barcode page (e.g. TCPDF example 027) at DPR 1 and 2: bars solid black and separated.
• Clip, blend mode, constant alpha, and soft masks still apply to snapped fills.

FemtoVG could get the same treatment separately.