[Chore] Install CPU torch wheels on Windows and Linux #138

The engine runs on CPU (device="cpu"), but pyproject.toml pulls CUDA-enabled torch on Windows (cu128 index) and Linux (PyPI default): multi-GB downloads with no use yet. Decided with the team lead: use the PyTorch CPU wheel index on Windows and Linux (OS markers), keep the PyPI default on macOS; CUDA comes back in M4 as a cuda extra. Also record in AGENTS.md that executable bits need git update-index --chmod=+x (the repo has core.fileMode=false).

Done when: uv lock resolves torch from the CPU index on Windows/Linux and from PyPI on macOS; CI passes on the CPU wheels.