QVAC-23799 feat[asr-ggml]: add opt-in CUDA GPU backend for whisper and parakeet

[Zbig9000]

🎯 What problem does this PR solve?

• asr-ggml could never run on CUDA. Both engines already resolved and reported CUDA at runtime — WhisperModel.cpp and ParakeetModel.cpp map ggml's CUDA* backend name to BackendId.CUDA (2), and gpu.test.js / parakeet-gpu-smoke.test.js already had cuda branches — but no build ever compiled the backend in.
• The cause was purely dependency wiring: vcpkg.json requested speech-cpp with vulkan (desktop/Android), metal (Apple) and opencl (Android), never cuda. On NVIDIA hosts a use_gpu / useGPU request therefore always landed on Vulkan.

📝 How does it solve it?

• New cuda feature in vcpkg.json forwarding to speech-cpp[cuda] → ggml-speech[cuda] (GGML_CUDA=ON, nvcc resolved by the port).
• New ASR_CUDA CMake option appending "cuda" to VCPKG_MANIFEST_FEATURES, following the SD_CUDA precedent in diffusion-cpp.
• New build:cuda / build:native:cuda scripts (bare-make generate -D ASR_CUDA=ON).
• Opt-in, not default, because it needs nvcc on the build host, which the prebuild runners do not carry. Only the NVIDIA driver is needed at runtime.
• CUDA is compiled alongside Vulkan rather than replacing it. ggml registers CUDA ahead of Vulkan, so a GPU request prefers CUDA and falls back to Vulkan when no supported device is present.
• Apple and Android are excluded via "supports": "!(osx | ios | android)".
• The cuda feature carries its own speech-cpp floor, 2026-08-24#1 — the version that introduced the feature. Feature dependencies only apply when the feature is selected, so default builds keep resolving at 2026-08-18 and are unaffected.
• No C++ changes.

No companion PRs needed

• Registry: speech-cpp[cuda] → ggml-speech[cuda] already shipped in tetherto/qvac-registry-vcpkg@cc90801 (QVAC-23802, speech-cpp 2026-08-24#1). Nothing to add there.
• vcpkg-configuration.json baseline: unchanged. The pinned baseline dca20a94 carries only speech-cpp 2026-08-10 / 2026-08-07, yet main already requires >= 2026-08-18 and builds green — version constraints resolve against fetched registry HEAD, not the baseline commit. 2026-08-24#1 resolves the same way.
• Version bump: none, matching the established flow — QVAC-23758 fix: drain queued streaming output before JobEnded in parakeet endStreaming #3950 added an Unreleased entry without touching version, and bumps land in dedicated release PRs (e.g. chore[mod]: release asr-ggml 0.3.1, ...). This PR adds to ## [Unreleased].

🧪 How was it tested?

• New scripts/__tests__/build-backends.test.js (6 tests) pins the wiring: the option defaults to OFF, build:native does not enable it, ASR_CUDA maps to the cuda manifest feature, the feature forwards to speech-cpp[cuda] with default-features: false, the platform guards match, and the CUDA floor is not below the base floor.
• node --test "scripts/__tests__/*.test.js" → 46/46 pass.
• Mutation-checked the two CMake regex assertions (option renamed, feature renamed, default flipped to ON) — each correctly fails, so the assertions are not vacuous.
• vcpkg format-manifest accepts the manifest. It also wants to canonically reorder the pre-existing dependency blocks, which the committed file does not follow, so that churn was left out to keep the diff reviewable.
• Prettier clean on the new test file.

Not covered: an actual -D ASR_CUDA=ON compile and a CUDA-vs-Vulkan runtime check. Both need a host with the CUDA toolkit; by design this PR adds no CI job, so the first real CUDA build is manual. The claim that ggml prefers CUDA over Vulkan is documented but unmeasured.

🔌 API Changes

No API changes. BackendId.CUDA (2) was already exported and already returned by getBackendInfo(); this PR only makes it reachable.

# opt into CUDA at build time
npm run build:cuda      # or: bare-make generate -D ASR_CUDA=ON

---

• To see the specific tasks where the Asana app for GitHub is being used, see below:
  • https://app.asana.com/0/0/1217592019459441

[github-actions]

Review Status

Current Status: ❌ PENDING
Approvals so far: none

Pending reviews: Needs 1 Management or Team Lead, and 1 more from Management, Team Lead, or Member.

[GustavoA1604]

Do not open from fork