Prototype Apple container sidecar torch frontend

[kevmo314]

Summary

• add a session-scoped lupine.sidecar(...) frontend that launches a Linux/arm64 PyTorch worker through Apple container on macOS
• add a custom CUDA PyTorch worker image that bundles the LUPINE libcuda/NVML shims
• add preflight handling for Apple container services, missing images, and missing container installs
• forward factory and tensor ops through a session-scoped TorchDispatchMode instead of per-op Python boilerplate

Validation

• uv run pytest
• uv run --python 3.13 --with pytest --with torch --with numpy --with-editable ./python pytest -q python/tests/test_lupine_adapter.py
• built and cached lupine-pytorch-worker:cuda-13.1.0 with Apple container
• proved native macOS Python frontend returns [0.0, 2.0, 4.0, 6.0, 8.0, 10.0, 12.0, 14.0] through sidecar CUDA PyTorch on an RTX 4090

Notes

This is intentionally a prototype backend. It proves the local macOS frontend plus Linux sidecar approach, but there are still CUDA backend limitations outside the Python forwarding layer.