feat(stablediffusion-ggml): build a ROCm variant

[localai-bot]

Description

stablediffusion-ggml had no ROCm build. The Makefile already carried a hipblas branch, but no row in .github/backend-matrix.yml built it and the gallery's amd: mapping was commented out. SystemState.Capability() therefore found no amd key for this backend and fell back to default, so AMD users silently got cpu-stablediffusion-ggml unless they installed the Vulkan variant by hand.

This PR:

• adds the hipblas row to the backend matrix (-gpu-rocm-hipblas-stablediffusion-ggml, rocm/dev-ubuntu-24.04:7.2.1, Dockerfile.golang), modelled on the existing whisper ROCm row rather than the llama-cpp one: builder-base-image only works with Dockerfile.llama-cpp, which has the builder-prebuilt target;
• adds the rocm-stablediffusion-ggml and rocm-stablediffusion-ggml-development gallery entries with their docker.io mirrors, and points amd: at them;
• drops -DGGML_HIPBLAS=ON from the Makefile. At the pinned sd.cpp commit, SD_HIPBLAS already does set(GGML_HIP ON) internally; GGML_HIPBLAS is the pre-rename ggml name and only produced an unused-variable warning;
• adds gfx1151 to the Makefile's local AMDGPU_TARGETS default so it matches the 11-target list the workflows pass in as a build arg.

Notes for Reviewers

Verification run on this branch:

• go test ./core/gallery/ passes; capability lookup resolves amd -> rocm-stablediffusion-ggml for both the release and development metas.
• FORCE_ALL=true bun run scripts/changed-backends.js picks the new row up into the single-arch build and merge matrices. Shards land at 78-79 entries, well under the 256-job limit.

Not verified: the hipblas compile itself. There is no ROCm toolchain on the machine this was written on, so CI is the first real build.

One thing worth a reviewer's judgement is build time. This backend still compiles four separate CPU variants (avx/avx2/avx512/fallback), each recompiling the full GPU backend. For reference, from recent master runs: -gpu-rocm-hipblas-llama-cpp takes 126 min and -gpu-nvidia-cuda-13-stablediffusion-ggml takes 91 min, so this row will probably sit in the 2-3 hour range. That is under the 6-hour job ceiling but makes it the slowest row this backend has. Trimming AMDGPU_TARGETS for this backend, or moving it to the single-build GGML_CPU_ALL_VARIANTS pattern that llama-cpp uses, would both bring it down; I left both out of scope here.

No docs change: nothing under docs/content/ enumerates per-vendor backend images, and no API, CLI flag or config key changed. The variant is selected automatically by capability.

Signed commits

• Yes, I signed my commits.
• Documentation updated (docs/content/) for user-facing changes, or not applicable