RenderTargetBitmap now records the visual tree into an SKPicture on
the UI thread (mirroring the regular rendering pipeline) and replays
it into a GRContext-backed surface during the CompositionTarget's next
render pass, reading the pixels back into the caller's buffer. The
GRContext is resolved from the draw canvas (SKCanvas.Context), so the
mechanism is platform-neutral and applies to every Skia backend; when
the target renders in software (raster canvas), pending render jobs
fail fast and the previous raster path is used as fallback.

Opting out is possible via the new
FeatureConfiguration.Rendering.UseGpuRenderTargetBitmap flag, since
GPU rasterization can produce slightly different anti-aliasing than
the software rasterizer.

RenderAsync now completes asynchronously on Skia (one render pass
later) instead of synchronously inside the call.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

GPU rendering of RenderTargetBitmap is now always on when the target
renders with a GRContext. When_CornerRadius_AntiAliasing consequently
can no longer pin itself to the software rasterizer, and no RMSE
threshold separates valid GPU-vs-CPU anti-aliasing variance (0.0325)
from a simulated non-anti-aliased render (0.0349) on this image, so
the test now asserts the invariant directly: a meaningful fraction of
the pixels in a ring around the circle's edge must blend between the
two colors (~34% for the CPU rasterizer, 0% for a jagged render). Its
software-rendered reference asset is removed with it.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Both overloads carried identical copy-pasted bodies differing only in
scaledSize and, through historical drift, in the fallback used when
element is null (PublicRootVisual vs RootElement). They now share one
core, unified on RootElement: the window's root visual, i.e. what is
presented on screen including popups.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The hardware and software paths shared no code despite doing the same
thing: both now record the SKPicture through one method (owning the
layout-clip removal and the IsSoftwareRenderer record-time flag) and
replay it through one method (surface, optional resampling, pixel
read-back). The only difference between the two paths is the surface
factory: GRContext-backed for hardware, raster for the fallback.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The drag visual was captured with RenderAsync, which now completes
during a later render pass; the resulting dispatcher yield broke the
synchronous DragStarting -> DragStarted -> DragEnter/DragOver sequence
(matching WinUI) that the drag-and-drop runtime tests assert, failing
them on all Skia CI targets. The capture now uses a synchronous
software render (its behavior before the GPU path was introduced),
while the public RenderAsync stays asynchronous.

Also hardens RenderAsync against overlapping calls on the same
instance: calls are serialized with an async gate and the pending GPU
replay roots the pixel buffer it writes to, so a concurrent
EnsureBuffer can neither resize nor free it mid-write.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>