QVAC-13445: Feature media generation

[aegioscy]

🎯 What problem does this PR solve?

• Adds a new @qvac/diffusion-cpp package (packages/lib-infer-diffusion) for text-to-image and image-to-image generation using stable-diffusion.cpp, built for the Bare Runtime
• Supports SD1.x, SD2.x, SDXL, SD3, and FLUX.2 [klein] model families
• GPU acceleration via Metal (macOS/iOS), Vulkan (Linux/Android/Windows), and OpenCL (Android/Adreno), with CPU fallback

📝 How does it solve it?

• C++ native addon wrapping stable-diffusion.cpp via N-API, following the same architecture as qvac-lib-infer-llamacpp-llm (single-job runner, runJob/activate/cancel/destroyInstance)
• vcpkg overlay ports for ggml and stable-diffusion-cpp with custom patches for runtime backend selection, abort callbacks, failure-path cleanup, and Android Vulkan diagnostics
• Android dynamic backend loading (GGML_BACKEND_DL) with libqvac-diffusion-ggml-* naming to avoid symbol conflicts with system ggml libraries; CPU backends remain statically linked (GGML_CPU_STATIC)
• JS layer (ImgStableDiffusion) extends BaseInference from @qvac/infer-base, providing load() → run(prompt, params) → cancel() lifecycle with progress callbacks and PNG output via QvacResponse
• CI/CD: prebuild workflows for Linux (x64/arm64), macOS (arm64/x64), iOS, Android, and Windows; integration test workflows with GPU runner matrix
• Documentation: README, build.md, architecture.md, CHANGELOG.md, NOTICE, examples, and TypeScript type declarations

Package structure

Layer	Files	Purpose
JS API	index.js, index.d.ts, addon.js	Public API, TypeScript types, native addon loader
C++ handlers	SdCtxHandlers, SdGenHandlers	Config parsing for context creation and generation params
C++ model	SdModel	Model lifecycle, load()/process()/cancel(), progress/abort callbacks
C++ utils	BackendSelection, LoggingMacros	GPU backend selection (Metal/Vulkan/OpenCL/CPU), logging
Build	CMakeLists.txt, vcpkg/	cmake-bare build, vcpkg overlay ports for ggml + stable-diffusion-cpp

Supported platforms

Platform	Architecture	GPU Backend
macOS	arm64, x64	Metal
iOS	arm64	Metal
Linux	arm64, x64	Vulkan
Android	arm64	Vulkan, OpenCL
Windows	x64	Vulkan

🧪 How was it tested?

• C++ unit tests (test/unit/): model loading, cancel context, generation handlers, backend selection, single-step inference, full generation
• JS integration tests (test/integration/): model loading, image generation (SD2.1, SD3, SDXL, FLUX.2), API behavior (run/cancel, busy error, cancel-then-rerun)
• CI: prebuild workflows on all target platforms, integration tests on GPU runners
• Manual: verified on macOS arm64 (Metal) with FLUX.2 klein and SD2.1 models

🔌 API Changes

const { ImgStableDiffusion } = require('@qvac/diffusion-cpp')

const model = new ImgStableDiffusion(
  { diskPath: '/path/to/model.gguf' },
  { threads: 8, device: 'gpu' }
)

await model.load()

const response = model.run({
  prompt: 'a photo of a cat',
  width: 512,
  height: 512,
  steps: 20,
  cfg_scale: 7.0,
  seed: 42
})

response.onUpdate((data) => {
  if (data instanceof Uint8Array) {
    // PNG image bytes
  } else if (typeof data === 'string') {
    // progress tick: {"step":5,"total":20,"elapsed_ms":1234}
  }
})

await response.await()

[gianni-cor]

/review

comments addressed