Drain in-flight graphics tasks before NativeEngine teardown

[bkaradzic-microsoft]

Problem

NativeEngine::Dispose only called m_cancellationSource->cancel(). arcana cancellation short-circuits work at scheduling boundaries, but a task body that is already running on a threadpool thread keeps going to completion.

The graphics-loading entry points (loadProgram, loadTexture, loadCubeTexture) schedule their work as threadpool tasks that capture a raw Graphics::Texture* / call into bgfx. On app teardown, if the suite finishes while one of these bodies is mid-flight, Dispose cancels and teardown proceeds to free the texture / tear down the bgfx context — and the still-running task then touches freed resources, causing an access violation:

Access violation. Exception Code c0000005
  2: Plugins/NativeEngine/Source/NativeEngine.cpp  LoadTextureFromImage
  3: arcana .../internal_task.h  output_wrapper<...>::invoke<...LoadTexture...lambda...>
  ...
  6: cppwinrt .../Windows.System.Threading.h  WorkItemHandler::Invoke

This reproduces at shutdown, after the test run reports it finished.

Fix

Add a small AsyncTaskTracker (mutex + condition_variable + counter) whose RAII Scope token is captured into the four graphics-touching threadpool lambdas (the loadProgram body, the loadTexture body, and the two LoadCubeTextureFromImages inline continuations). The CPU-only image-parse bodies are intentionally not tracked.

Dispose now cancels and then waits for those tokens to drain before returning, so teardown can't free graphics resources out from under a running task. The token decrements on the threadpool thread (never on a JS-thread .then continuation), so Dispose — which runs on the JS thread — can never deadlock waiting on its own continuations. Counting is bound to the token's lifetime, so the increment/decrement pair can't be split by an exception.

A defensive if (cancellationSource->cancelled()) return; early-out is also added at the top of the two threadpool load bodies, so once teardown has begun they skip touching graphics resources entirely.

Testing

• NativeEngine and Playground build clean (RelWithDebInfo, x64, D3D11).
• Ran the previously-crashing Sprites test (loadTexture path) and a cubemap test (LoadCubeTextureFromImages path) headless — both exercise the load paths and tear down promptly with no hang and no crash.

[Copilot]

Pull request overview

This PR hardens NativeEngine shutdown by ensuring any in-flight threadpool work that touches bgfx/graphics resources has completed before teardown proceeds, preventing use-after-free crashes during app/test shutdown.

Changes:

• Added an AsyncTaskTracker (counter + condition_variable) and TrackAsyncTask() RAII token to track graphics-touching async work.
• Updated NativeEngine::Dispose() to cancel work and then block until tracked tasks drain.
• Captured tracking tokens into loadProgram, loadTexture, and cube-texture inline continuations; added cancellation early-outs for the two threadpool bodies.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
Plugins/NativeEngine/Source/NativeEngine.h	Introduces AsyncTaskTracker + TrackAsyncTask() and stores the tracker on NativeEngine.
Plugins/NativeEngine/Source/NativeEngine.cpp	Implements tracker logic, waits during Dispose(), and wires tracking/cancellation checks into async graphics load paths.