refactor: harden CUDA backend and solver contracts

[KemengHuang]

Summary

• harden CUDA backend loading, deterministic system topology, component boundaries, and reusable buffer/CUB workspaces
• add scene-config contracts, test sharding, benchmark entrypoints, ADRs, and performance evidence
• correct thin-shell reference weighting and full-thickness stretch/bending conventions
• enable semi-implicit Newton termination by default with K_min = 6
• synchronize the cloth-sample submodule and replace unsafe QR-SVD sign-copy calls with typed CUDA-safe arithmetic

Validation

• Repository Contracts: 31/31
• Core: 36 cases / 1005 assertions
• CUDA sim_case: 95 cases / 14212 assertions
• full XMake CUDA build and device-link: passed
• full CMake CUDA build and device-link: passed
• CUDA float Wilkinson-shift smoke on sm_120: passed
• MkDocs + Doxygen build: passed