fix(llama-cpp): preserve GPU layers during option passthrough

[localai-org-maint-bot]

Description

Fixes #11191.

The latest llama.cpp argument parser asserts that main and draft n_gpu_layers still contain negative initialization sentinels when parser options are registered. LocalAI resolves those fields from the model request before applying generic passthrough flags, so any passthrough option could abort the backend and surface as an EOF from the frontend.

This stages the parser sentinels while applying raw flags, then restores LocalAI's resolved GPU-layer values unless the passthrough explicitly supplies a main or draft GPU-layer flag. Explicit auto remains authoritative.

Notes for Reviewers

• Added standalone regression coverage for staging/restoration, numeric overrides, and explicit auto.
• Verified with NLOHMANN_INCLUDE=backend/cpp/llama-cpp/llama.cpp/vendor bash backend/cpp/run-unit-tests.sh (3 test files).
• Verified the backend preparation path copies the helper and test into the generated grpc-server tree.
• Full backend compilation was not run because it is a long build; CI covers the backend matrix.

Signed commits

• Yes, I signed my commits.
• Documentation updated (docs/content/) for user-facing changes, or not applicable

[mudler]

@localai-org-maint-bot did you reproduced it locally?

[localai-org-maint-bot]

Yes, but at the focused parser-path level rather than through a full LocalAI backend build. I reproduced the exact llama.cpp assertion locally by exercising the option-registration path with pre-resolved n_gpu_layers, then verified the staging/restoration fix with the standalone C++ regression suite. I did not compile and launch the complete llama-cpp backend end to end; the PR description calls that out explicitly.