[Velli20]

Fixes #498

Problem

Axial and radial shadings were sampled into 16 evenly spaced color stops. gradient_many_stops.pdf uses a stitching function with 199 linear segments, so Safe-PDF drew about three smooth bands where PDFium draws 40 sharp ones.

Change

• Function::breakpoints() returns a stitching function's /Bounds (empty for other types).
• ColorStops::from_function_domain keeps the 16-point even grid and adds two stops at each bound inside the shading domain: the left segment's end (sampled just below the bound) and the right segment's start. Both share a position, so each handover is a hard edge and linear segments are exact. Out-of-order bounds are skipped to keep positions sorted. Shadings without stitching get the same stops as before.
• interpolate (the web canvas CPU rasterizer) finds the segment by binary search instead of a linear scan, now that a gradient can carry hundreds of stops.

Conformance

Mismatch on the case drops from 89% to 63% and the band pattern matches PDFium. cargo conformance verify conf2-1ca6fb38db80 --allow-different exits 0. The case now fails differently (offset), which is expected: PDFium samples shading functions into a 256-entry table at t = i/256 with a floor lookup, which shifts the bands by about 3 px and never reaches t = 1 (hence its violet rather than red extended end). Safe-PDF keeps exact sampling.

No new unit test was added, per the project's testing preference. The existing pdf-function, pdf-shading and pdf-canvas tests pass, and clippy is clean.

🤖 Generated with Claude Code