fix(backends): enable ROCm/HIP GPU offload for ggml audio backends (#10666)

[localai-bot]

Summary

Fixes #10666. The ROCm variants of the ggml-based audio backends — rocm-qwen3-tts-cpp, rocm-omnivoice-cpp, and the same-family acestep-cpp / vibevoice-cpp — silently ran on CPU instead of the AMD GPU. The reporter's log showed [Load] Talker backend: CPU (CPU threads: 4); upstream qwentts.cpp selects the best device via ggml_backend_init_best(), so a CPU result means the HIP backend was never compiled into the binary.

Two coupled build defects, both invisible without a ROCm build:

1. Wrong CMake flag (primary). The Makefiles passed -DGGML_HIPBLAS=ON, but the vendored ggml only understands -DGGML_HIP=ON (GGML_HIPBLAS was removed upstream — no references anywhere in the source tree). The flag was a silent no-op: the ggml-hip target was never created and no GPU code was built. The blocks also lacked the ROCm compiler (CC/CXX) and AMDGPU_TARGETS a HIP build requires.
2. Missing link (secondary). The CMake foreach that links ggml GPU backends into the module listed blas cuda metal vulkan [sycl] but not hip — so even a built ggml-hip would not be linked and its static backend-registration constructor would never run.

CUDA users were unaffected because cublas passes the correct GGML_CUDA=ON and the foreach already links cuda — which is exactly why the two code paths diverged.

Changes

Across qwen3-tts-cpp, omnivoice-cpp, acestep-cpp, vibevoice-cpp (Makefile + CMakeLists):

• Replace the no-op GGML_HIPBLAS block with the proven llama-cpp hipblas recipe: ROCm clang CC/CXX, AMDGPU_TARGETS (default list includes gfx1201, the reporter's R9700), and -DGGML_HIP=ON.
• Add hip to each linking foreach so ggml-hip and its propagated ROCm deps (hip::host, roc::rocblas, roc::hipblas) land on the module's link line.

No run.sh change is needed: the HIP backend is statically linked into each CPU-SIMD variant .so and auto-selected at load time.

Testing

• All four Makefiles parse with BUILD_TYPE=hipblas and resolve to -DGGML_HIP=ON -DAMDGPU_TARGETS=….
• Confirmed the vendored ggml defines GGML_HIP and creates target ggml-hip; PRIVATE deps of a static lib propagate to consumers, so linking ggml-hip pulls the ROCm runtime automatically.
• The *_HAVE_HIP compile-defs newly set on acestep/vibevoice are inert (defined but unreferenced) — no behavioral change beyond enabling GPU.

⚠️ I do not have a ROCm toolchain/GPU locally, so final validation is the ROCm CI build plus confirmation on the reporter's gfx1201 box.

Assisted-by: Claude:claude-opus-4-8[1m] [Claude Code]