ci: build portable llama.cpp release binaries

[LauraGPT]

Summary

• disable ggml native CPU tuning in the llama.cpp release workflow
• explicitly disable AVX/AVX2/AVX512/VNNI/FMA/F16C/BMI2 for conservative prebuilt x64 assets
• document why optimized builds should be separate release assets

Why

The v0.1.2 Windows prebuilt binaries can crash with SIGILL / 0xC000001D on user machines because ggml defaults to native CPU optimization on CI runners. This makes the generic x64 release artifact depend on the runner CPU instruction set.

Fixes #3036. Also related to FunAudioLLM/SenseVoice#307.

Verification

• git diff --check
• parsed .github/workflows/build-llamacpp-binaries.yml with PyYAML
• configured runtime/llama.cpp with the release flags and local pinned llama.cpp source
• built llama-funasr-vad successfully with the conservative flags

[gemini-code-assist]

Note

Gemini is unable to generate a review for this pull request due to the file types involved not being currently supported.