Add llama-cpp-python backend support

[Copilot]

Adds llama-cpp-python as a supported inference backend, consistent with the existing vllm and vllm_offline integrations and adapted from the outlines llamacpp model.

Changes

• src/gimkit/models/llamacpp.py: New LlamaCpp subclass overriding __call__ with GIMKit input/output handling; _ensure_response_suffix injects RESPONSE_SUFFIX as a stop token; from_llamacpp factory function
• src/gimkit/models/__init__.py / src/gimkit/__init__.py: Export from_llamacpp
• pyproject.toml: New llamacpp optional dependency group (llama-cpp-python>=0.3.0)
• tests/models/test_llamacpp.py: Tests for factory, call dispatch, error handling, and stop-token injection; uses an autouse fixture patching LlamaCppTokenizer so the suite runs without llama-cpp-python installed

Usage

from llama_cpp import Llama
from gimkit import from_llamacpp, guide

llm = Llama(model_path="model.gguf")
model = from_llamacpp(llm)

result = model(f"The capital of France is {guide(name='city')}.")
print(result.tags["city"].content)  # "Paris"

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.