qvac-fabric: fix BUILD_LLAMA=OFF configure failure (port-version 1)

[gianni-cor]

Summary

Fixes a portfile bug introduced together with the new `gpu-backends`/`llama` feature split (#133) and the `8189.0.0` bump (#135): when a consumer requests `qvac-fabric` without the `llama` feature, the configure step aborts with:

```
-- BUILD_LLAMA is OFF: skipping LLAMA targets.
-- Generating embedded license file for target: common
CMake Error at cmake/license.cmake:38 (message):
Target 'common' does not exist
```

Root cause: the portfile passes `-DLLAMA_MTMD=ON` unconditionally. `LLAMA_MTMD` transitively enables `LLAMA_BUILD_COMMON`, which causes upstream's `CMakeLists.txt` to call `license_generate(common)` — but `BUILD_LLAMA=OFF` (which is what `vcpkg_check_features` derives when `llama` is not requested) skips defining the `common` target.

Fix: move the `LLAMA_*` configure flags into a feature-gated `LLAMA_OPTIONS` list. With `llama` on, behaviour is unchanged. With `llama` off, pass `-DLLAMA_MTMD=OFF` + `-DLLAMA_BUILD_COMMON=OFF` so the license-generation block is skipped.

Test plan

• Reproduced the failure with tetherto/qvac#1874's `qvac-lib-infer-nmtcpp` (`qvac-fabric` declared as `default-features: false` + `features: ["gpu-backends"]`) cross-building for `arm64-android` from a macOS host.
• After this PR merges, re-run the same build and confirm `qvac-fabric[core,gpu-backends]:arm64-android@8189.0.0#1` configures and installs successfully.
• Confirm `qvac-fabric[core,gpu-backends,llama]` (LLM/Embed consumers, the previous behaviour) still builds — the `llama` branch of the new conditional emits the same flags as before, so this is a no-op for them.

Made with Cursor

[gianni-cor]

Validated locally by overlaying this fixed port (`VCPKG_OVERLAY_PORTS`) into the tetherto/qvac#1874 build of `qvac-lib-infer-nmtcpp` (`qvac-fabric` declared with `default-features: false` + `features: ["gpu-backends"]`) cross-compiling for `arm64-android` from a macOS-arm64 host:

```
Installing 5/7 qvac-fabric[core,gpu-backends]:arm64-android@8189.0.0#1...
…
All requested installations completed successfully in: 3.4 min
…
[14/14] Linking CXX shared library qvac__translation-nmtcpp@3.bare
```

Configure now finishes (no more `Target 'common' does not exist` error) and the addon builds end-to-end.

Spot-check that the `llama` branch is still fine: previous validation in #135 built `qvac-lib-infer-llamacpp-llm` (`features: [llama, gpu-backends]` implicitly) for `arm64-osx` against `8189.0.0#0`. The new `#1` portfile preserves that branch byte-for-byte (same `-DLLAMA_*` flag set), so no regression risk for LLM/Embed consumers.

[gianni-cor]

Force-pushed `5b9dd7c`: per review feedback, hoist all the always-OFF `LLAMA_*` flags (`LLAMA_CURL`, `LLAMA_BUILD_TESTS`, `LLAMA_BUILD_TOOLS`, `LLAMA_BUILD_EXAMPLES`, `LLAMA_BUILD_SERVER`, `LLAMA_ALL_WARNINGS`) into the `LLAMA_OPTIONS` initialization so they're set in both branches.

Original commit only set those flags in the `llama` branch, on the assumption that `BUILD_LLAMA=OFF` already prunes the upstream code paths they affect. That's true today (`-- BUILD_LLAMA is OFF: skipping LLAMA targets`) but defensive symmetry is cheaper than a future regression. Net structure:

```cmake
set(LLAMA_OPTIONS
-DLLAMA_CURL=OFF
-DLLAMA_BUILD_TESTS=OFF
-DLLAMA_BUILD_TOOLS=OFF
-DLLAMA_BUILD_EXAMPLES=OFF
-DLLAMA_BUILD_SERVER=OFF
-DLLAMA_ALL_WARNINGS=OFF
)
if("llama" IN_LIST FEATURES)
list(APPEND LLAMA_OPTIONS -DLLAMA_MTMD=ON)
else()
list(APPEND LLAMA_OPTIONS
-DLLAMA_MTMD=OFF
-DLLAMA_BUILD_COMMON=OFF
)
endif()
```

Re-validated: nmtcpp arm64-android cross-build still configures (both `arm64-android-dbg` and `-rel`) with this revision. Behaviour for the `llama` branch is identical to before.