Detect CUDA in sweep skills and exercise GPU paths when available

[brendancol]

Summary

• Each sweep skill (sweep-accuracy, sweep-api-consistency, sweep-metadata, sweep-performance, sweep-security, sweep-test-coverage) now probes the host with numba.cuda.is_available() before dispatching subagents and passes a CUDA_AVAILABLE flag into the agent prompt.
• On a CUDA host, subagents run cupy and dask+cupy backends end-to-end during audit and /rockout verification. Each sweep gets a tailored block — accuracy runs the kernel on small inputs, performance times the GPU path, security tests adversarial shapes, metadata checks attrs/coords/dims on cupy DataArrays, api-consistency smoke-tests cupy entry points, and test-coverage requires new cupy tests to execute locally.
• On a CPU-only host, subagents fall back to static review and tag the state CSV's notes column with cuda-unavailable so a future re-run on a GPU host knows to re-validate the GPU paths.

Test plan

• Run /sweep-performance --top 1 on a CUDA host; confirm the dispatched subagent receives CUDA_AVAILABLE=true and exercises the cupy path
• Run the same sweep on a CPU-only host; confirm the state CSV row carries the cuda-unavailable note
• Spot-check that the rendered prompt for each of the six sweeps contains the CUDA instruction block