fix: respect GGML_METAL=OFF on macOS

[provoke210]

Fixes #2050.

On macOS, -DGGML_METAL=OFF is overwritten by the Apple configuration block: the CMake cache ends up with GGML_METAL:BOOL=ON and the Metal backend is still built. Remove FORCE from that cache assignment so an explicit user setting is preserved while Metal remains enabled by default on both Apple architectures.

The embedded Metal library setting is unchanged. This also adds an Unreleased changelog entry.

Validation

Tested on Apple M5 / macOS 26.5 (arm64), Python 3.12.14, CMake 4.4.3, against main d73664636c2e3b25ce5949a9b85bba690684cdf3 and its unchanged llama.cpp submodule 5266f24da75dc449bd56cbed7addb9c8e4a6a73e.

• Before the fix, configuring with -DGGML_METAL=OFF produces GGML_METAL:BOOL=ON and a ggml-metal target.
• After the fix, fresh default and explicit-ON configurations keep Metal enabled; a fresh explicit-OFF configuration keeps it disabled and has no Metal target. Reconfiguring an existing build ON → OFF → ON also updates the cache and target graph correctly.
• Built a fresh wheel with CMAKE_ARGS="-DGGML_METAL=OFF" CMAKE_BUILD_PARALLEL_LEVEL=6 python -m build --wheel --no-isolation, then installed it into the test environment. An isolated import confirms the installed wheel is used, contains no Metal library, and reports llama_supports_gpu_offload() == False.
• HF_HUB_OFFLINE=1 python -I -m pytest -q: 26 passed, including the existing generation, embedding, and recurrent/hybrid cache tests with their downloaded model fixtures.
• python -m ruff check llama_cpp tests, python -m ruff format --check llama_cpp tests, and git diff --check: passed.

The Metal-enabled configurations were configured but not compiled. Intel macOS, Linux, and Windows execution were not tested locally.

AI assistance: OpenAI Codex assisted with implementation, review, local validation, and this PR description.