cuda version check based on hardware

[imprsnst]

Summary

Launchers choose the index from the card. webui.sh / webui.bat read
nvidia-smi --query-gpu=compute_cap, take the highest across GPUs, and pick cu130 for Blackwell
or cu126 for everything older. cu126 is the last index still built for Maxwell..Volta, so nothing
old gets stranded, and it moves every Windows user off the frozen 2.6.0 at the same time. A driver
too old falls back to cu126.

The install now also passes --no-sources-package torch, which is the piece that actually lets the
detected index beat the pyproject pin.

An escape hatch --torch-index (and
INLINE_TORCH_INDEX) takes a short name, a full index URL, or cpu. Blackwell users whose driver
predates CUDA 13 want --torch-index cu128.

A post-install check fails loudly if a CPU-only wheel slipped through, which can still happen if
PyPI ever outranks the CUDA index on version.

core/pyproject.toml is retargeted to cu126 and uv.lock regenerated, so the bare uv sync path
on Windows moves off 2.6.0 too.

Verified on Blackwell hardware

RTX PRO 4500 Blackwell, sm_120, driver 595.71.05 / CUDA 13.2, 32 GB

Closes #28

Checklist

• npm run typecheck && npm run lint && npm run test pass
• cd core && ruff check . && uv run pytest -q pass (if the engine changed)
• Commits use Conventional Commits and are signed off (git commit -s)
• Change is small and focused, and matches the surrounding style