Summary

• patch the pinned llama.cpp UI CMake to treat single-arch macOS builds targeting a non-host architecture as a host-tool case for llama-ui-embed
• keep the generated llama-ui library and llama-server objects on the requested CMAKE_OSX_ARCHITECTURES slice
• allow Darwin amd64 builds on Apple Silicon hosts without requiring Rosetta just to run the build-time UI embedder

Fixes #16500.

Testing

• cmake -DPATCH_DIR=/tmp/ollama-16500/llama/compat -P /tmp/ollama-16500/llama/compat/apply-patch.cmake
• cmake -S llama/server -B build/test-llama-server-amd64-ui-host ... -DCMAKE_OSX_ARCHITECTURES=x86_64 -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0
• cmake --build build/test-llama-server-amd64-ui-host --target llama-ui-assets --parallel 4
• file build/test-llama-server-amd64-ui-host/_deps/llama_cpp-build/tools/ui/llama-ui-embed-host

The generated rules build llama-ui-embed-host without -arch x86_64, while llama-ui still compiles with -arch x86_64. The helper was reported as Mach-O 64-bit executable arm64 and llama-ui-assets completed.