Fix release deployment: compile acestep.cpp on the user's machine for real GPU acceleration

[Copilot]

Release bundles were shipping CPU-only binaries compiled on GPU-less GitHub runners, meaning no user ever got GPU acceleration regardless of their hardware.

Approach

Shift the acestep.cpp compilation to first launch on the user's machine — same pattern as node-gyp native addons — where the actual GPU is present and detectable.

New files

build.sh / build.bat — clone acestep.cpp if absent, auto-detect GPU, compile with correct flags:

./build.sh          # auto-detect (CUDA / ROCm / Vulkan / Metal / OpenBLAS)
./build.sh --cuda   # force CUDA
./build.sh --rocm   # force ROCm/HIP
./build.sh --cpu    # CPU-only

Detection matrix:

Hardware	cmake flag
NVIDIA	-DGGML_CUDA=ON
AMD	-DGGML_HIP=ON
Vulkan	-DGGML_VULKAN=ON
macOS	Metal auto-detected
OpenBLAS	-DGGML_BLAS=ON (coexists with GPU)

release.yml

• Removed cmake compilation from CI jobs entirely — no GPU runners needed, all four platform bundles now assemble on ubuntu-latest
• Bundle ships build.sh, build.bat, models.sh, models.bat
• Generated start.sh / start.bat check on each launch: no binaries → run build.sh; no *.gguf → run models.sh; then start the server
• Added windows-x64 platform (.zip archive)
• Fixed ACESTEP_CPP_REPO to canonical audiohacking/acestep.cpp

ci.yml

• Replaced build-acestep-cpp job (which tested a manual cmake build no user ever runs) with test-build-script — runs bash build.sh --cpu end-to-end and verifies binaries land in bin/
• Cache keyed on build.sh hash + repo URL so reruns are fast

setup.sh

Calls build.sh automatically instead of printing manual cmake instructions.

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.