Ship whisper-gpu (Windows) as a C-API DLL — cross-build with MinGW, drop the Windows-box dependency

Background

PR #4338 (RFC #4333, landed on main as 4d7f00a6) ships the Windows ASR GPU backend (whisper Vulkan) as prebuilt MSVC static libs hosted on the whisper-gpu-<ver> release tag (currently whisper-gpu-1.9.1), consumed via -DASR_USE_PREBUILT_WHISPER_GPU=ON. This was necessary because MSVC compiles ggml-vulkan.cpp too slowly to finish inside GitHub's 6-hour job timeout on the stock windows-latest runner.

Problem

The prebuilt libs are MSVC-ABI static libs, so regenerating them — e.g. on any whisper.cpp bump — requires a native Windows box (scripts/build/build-whisper-vulkan-windows.ps1). They can't be cross-compiled from Linux with MinGW because MinGW and MSVC have incompatible C++ ABIs and static linking needs whole-library ABI compatibility. That's a recurring maintenance friction point: every whisper bump means booting into Windows to rebuild and re-upload the asset.

Proposal

Repackage whisper-gpu as a C-API DLL, exactly the way sherpa-onnx is already shipped. A C-ABI boundary (exported C functions, self-contained runtime) is stable across toolchains — which is why the MSVC-built app already links sherpa's DLL fine. whisper.h is a C API, so this is viable.

Once it's a C-API DLL:

• The MSVC app links the DLL's import lib and loads it at runtime (no C++ across the boundary).
• The DLL itself can be built with MinGW-w64, cross-compiled on Linux (the toolchain is already installed on the dev box) — no Windows machine required.
• The asset can then be regenerated in CI on a fast Linux runner, replacing the manual Windows build.

Tasks

• Build whisper + ggml + ggml-vulkan as a shared library exporting the whisper.h C API (BUILD_SHARED_LIBS, -static-libgcc -static-libstdc++ so the DLL is self-contained).
• Cross-compile for win-x64 via x86_64-w64-mingw32 on Linux; produce the import lib (.dll.a / .lib) the MSVC app links against.
• Sort out the Vulkan loader for the MinGW build (link vulkan-1, ship vulkan-1.dll — already handled app-side).
• Update the ASR_USE_PREBUILT_WHISPER_GPU CMake path to link the DLL import lib instead of the static libs, and deploy the DLL alongside the exe (mirror the sherpa-onnx deploy).
• New CI workflow to build + publish the whisper-gpu C-API DLL asset from Linux (retire the manual Windows producer script).
• Runtime smoke test: confirm the Vulkan GPU path works with the MinGW-built DLL loaded into the MSVC-built AetherSDR.exe.

Notes

• This is a maintenance / infra improvement, not user-facing — the current MSVC-static-lib approach works and ships GPU today.
• Model to follow: the existing sherpa-onnx C-API DLL packaging (scripts/setup/setup-sherpa-onnx.*), which already bridges the MinGW/MSVC toolchain gap.
• Related: PR #4338, scripts/build/build-whisper-vulkan-windows.ps1 (current Windows producer).

Start by reading scripts/build/build-whisper-vulkan-windows.ps1 and the existing scripts/setup/setup-sherpa-onnx.* packaging flow, then inspect the ASR_USE_PREBUILT_WHISPER_GPU CMake path. Compare the current static-library producer and deployment steps with the sherpa-onnx model. Done means a Linux CI workflow publishes the C-API DLL, the MSVC app deploys and loads it, and the Vulkan GPU smoke test passes.