Make optional deps safe, fix imports, harden llama-cli and runner registry

[harvatechs]

Motivation

• Prevent runtime/test failures when optional native deps (llama_cpp, torch) are not installed by treating them as optional and failing gracefully.
• Improve robustness of the FastAPI runner so model registry loading is independent of process CWD while still honoring ARIV_MODELS_YAML.
• Surface useful errors for llama.cpp CLI failures and avoid silent subprocess failures during streaming.

Description

• Treat llama_cpp and torch as optional runtime dependencies by guarding imports and branching behaviour when they are absent, and add an explicit ModelLoadError when attempting to load models without llama_cpp.
• Replace fragile parent-relative imports with absolute imports for the tools/config modules in core modules to avoid import-time crashes.
• Harden LlamaCLI.stream_chat to asynchronously collect stderr and raise a detailed RuntimeError (binary, model, exit code, stderr tail) if the subprocess exits with a non-zero status.
• Make runner registry path resolution independent of the current working directory by resolving ARIV_MODELS_YAML or falling back to a path relative to the package file.
• Update and add tests to reflect API/behavioural adjustments (VRAM manager API alignment, optional-torch handling, runner registry cwd-independence, and llama-cli non-zero-exit handling).

Testing

• Ran the full test suite with pytest -q and verified success: 30 passed, 2 skipped.
• Executed focused test runs including tests/test_runner_app.py, tests/test_llama_cli.py, tests/test_vram.py, tests/test_pipeline.py, and tests/test_orchestrator.py to validate changes, all of which passed in targeted runs.
• Skipped tests are CUDA-dependent and will be skipped automatically when CUDA is not available in the environment.

---

Codex Task