…PU edge accuracy

- CalcArc and stroke_calc_arc take their vertex count from the guarded
  GetArcSubdivisionCount and stroke_arc_subdivision_count used by round caps.
- StrokeOptions.ArcDetailScale rejects values not greater than zero.
- DefaultRasterizer: band boundary X is interpolated from the segment
  endpoint and rounded to nearest, coverage uses the exact doubled area,
  and the DDA walkers round row and column boundaries to nearest.
- Tests: ArcDetailScale validation, subdivision count monotonicity, and
  CPU/GPU parity theories for round joins at scales 0.01 to 16.
- Reference images updated for the corrected edge coverage.

Round join tessellation and CPU edge accuracy

Protect cache metadata, publish complete glyph entries, and lease reusable working buffers to each drawing operation. Keep outline construction and rasterization unchanged.

Add tests for concurrent drawing, outline reuse, complete publication, eviction, clearing, and buffer ownership. Update Fonts to 3.1.2 and ImageSharp to 4.1.2.

Validation: Full Release suite passed on .NET 8 and .NET 10, with 2,754 passed and 9 skipped per target.

Use `Task.Factory.StartNew` with `TaskCreationOptions.LongRunning` in the concurrent drawing text cache tests so each barrier participant gets a dedicated thread. This avoids thread-pool starvation and intermittent failures on small CI runners when other tests are also scheduling work.

Make DrawingTextCache thread-safe

The Paint image processing extension created a private text cache for
each frame canvas. The new overload passes an application-owned cache to
every frame canvas so glyph geometry is shared across draws, matching
the CreateCanvas overloads.

Cached color glyph layers re-create their paint brushes on replay. The
cache stored the device-space glyph origin and appended the device-space
delta after the drawing transform, so a draw whose transform carries an
origin applied that offset twice and the gradient drifted off the glyph.
Store the glyph origin before the drawing transform and apply the delta
before it.

Fix cached color glyph paint replay and add a Paint overload with a text cache

Transforming a gradient brush projected its geometry point by point,
which reduces any affine transform to a similarity. A COLRv1 radial
paint squashed by its PaintTransform into a flat ellipse became a
circle, and the pad region of that circle painted the toned thumbs-up
opaque dark.

GradientBrush carries a GradientTransform from the gradient's space to
the drawing, every brush has a constructor overload that takes it, and
Transform() composes it exactly for affine matrices. The old point
projection remains for projective matrices, which have no affine
inverse. Renderers fold the inverse: the linear brush into its axis,
the two-circle radial brush into its canonical transform, and the
others map each sample once when a transform is set. The glyph
renderer builds paint-space brushes with the paint's transform
composed after the glyph's.

Interpolation is unchanged: premultiplied, as the CSS Color 4
gradient rules require.

Every gradient payload starts with the six words of the inverse affine
transform. The draw tag's scene-size field widens to four bits to hold
the larger payloads. draw_leaf reads the transform once and composes
it per gradient kind: the linear brush folds it into its line
equation, the radial and elliptic brushes multiply it into their
canonical transforms, and the sweep brush carries the matrix and
translation.

Adds a CPU and GPU parity test over the four transformed brushes.

…-3.1

Compose gradient transforms on the brush and consume Fonts 3.1.3