[codex] Use cuGraph for Leiden clustering

[maarten-devries]

Summary

• Replace the CPU igraph Leiden implementation with a cuGraph-backed GPU implementation.
• Convert sparse neighbor connectivities into a cuDF edge list and run cugraph.leiden with the existing resolution and seed parameters.
• Add CUDA 12 RAPIDS graph dependencies and update docs/tests for the backend change.

Notes

The old n_jobs resolution sweep used CPU joblib workers. The GPU backend now runs the resolution sweep sequentially and warns when n_jobs is set, avoiding multiple Python workers competing for the same GPU.

Validation

• uvx ruff format --check src/scib_rapids/metrics/_nmi_ari.py tests/test_metrics.py tests/test_cross_validate.py
• uvx ruff check src/scib_rapids/metrics/_nmi_ari.py tests/test_metrics.py tests/test_cross_validate.py
• git diff --check
• python3 -m py_compile src/scib_rapids/metrics/_nmi_ari.py tests/test_metrics.py tests/test_cross_validate.py

Targeted pytest could not run on this local machine because the local Python environment does not have CUDA RAPIDS packages installed (cupy, cudf, and cugraph were unavailable).