build(cuda): remove system Toolkit runtime dependencies

[KemengHuang]

Summary:

• replace the remaining cuBLAS dot/norm calls with persistent raw-CUDA/CUB reductions, including stable scaled-square norms
• remove unused CUDA Toolkit library links from CMake and XMake and audit every wheel for forbidden dynamic Toolkit dependencies
• update compatibility diagnostics, packaging policy, install docs, and agent documentation for driver-only wheel deployment

Validation:

• CMake and XMake CUDA 13.2 backend builds passed; both DLLs have no CUDA Toolkit imports
• focused reductions: 12 assertions; CUDA backend: 274 assertions; simulation suite: 14,212 assertions in 95 cases
• 43 repository tests, 79 portable Python tests, full Doxygen/MkDoxy/MkDocs build, CUDA doctor probe, and four representative benchmarks passed