ci(linux): build fat package with GGML_BACKEND_DL + GGML_CPU_ALL_VARIANTS

[Geramy]

What

Switch the Linux x86_64 `ubuntu-latest-cmake` (cpu) and `ubuntu-latest-rocm` (HIP) builds from the AVX2/FMA/F16C portable baseline (#3) to a fat-package build:

```
-DGGML_NATIVE=OFF
-DGGML_BACKEND_DL=ON
-DGGML_CPU_ALL_VARIANTS=ON
```

The build now produces:

• `libstable-diffusion.so` — main shared library
• `libggml-cpu-sandybridge.so` — AVX
• `libggml-cpu-haswell.so` — AVX2 + FMA + F16C
• `libggml-cpu-skylakex.so` — + AVX-512F
• `libggml-cpu-icelake.so` — + AVX-512 VNNI
• `libggml-cpu-alderlake.so` — + AVX-VNNI + DOTPROD
• `libggml-cpu-x64.so` — no-SIMD fallback

At runtime, `ggml_backend_load_all_from_path` (already wired by upstream PR leejet#1448) dlopens each variant, queries `__builtin_cpu_supports`, and picks the highest-tier match. Same zip works on a 2014 Sandy Bridge laptop and a 2024 Alder Lake server — without the runner-of-the-day AVX-512 lottery that crashed master-593.

Tradeoff

Linux x86_64 zip: ~12 MB → ~50–80 MB. Acceptable IMO — Lemonade and similar consumers cache the extracted dir across model loads, so this is paid once at install. In exchange you stop choosing between portability and AVX-512 perf — you get both.

Why not Windows / macOS

• Windows AVX2 build already pins `GGML_NATIVE=OFF -DGGML_AVX2=ON`. Could get the same fat-package treatment for symmetry, but that's a separate change and not required to fix the consumer-side AVX-512 SIGILL pattern.
• macOS arm64 all-Apple-Silicon parts share a uniform NEON+DOTPROD+i8mm+bf16 baseline (M1+), so `-march=native` does not introduce a portability problem there. Current Metal flags are already optimal.

Verification plan

• Trigger a workflow_dispatch release on this branch (`create_release: true`).
• Inspect the resulting `sd-{hash}-bin-Linux-Ubuntu-24.04-x86_64.zip` — it should contain `libstable-diffusion.so` plus several `libggml-cpu-*.so` files.
• On an AVX-512-less host: `./sd-server -m model.safetensors` → uses haswell variant, no SIGILL.
• On an AVX-512 host: same command → uses skylakex/icelake/alderlake variant, perf matches a native-compiled build.
• Bump `sd-cpp` pins on lemonade-sdk/lemonade PR #1777 to the new tag and confirm `Test ollama (ubuntu-latest)` passes.

Lineage

Replaces #3's portable AVX2 baseline. #3 fixed the SIGILL but left AVX-512-class hosts running AVX2 code; this gets full perf on those hosts.