Align and minimize cudarc feature contracts

[shinaoka]

Summary

• align the normal CUDA dependency with tenferro's explicit CUDA 12.8 and dynamic-loading contract
• remove build-system CUDA detection and fallback loading/version features
• keep the build dependency role-minimal (std, driver, dynamic-loading, cuda-12080)
• preserve runtime, NVRTC, and NCCL on the normal dependency
• add contract tests for both dependency roles

Why

Cargo always builds normal and build dependencies as separate units. Giving both units the full normal feature set did not merge them and increased a controlled downstream release target by about 11 MiB. The build script only reads cudarc::driver::sys::CUDA_VERSION, so its dependency should share the explicit CUDA floor/loading mode without inheriting runtime, NVRTC, or NCCL.

The final role-minimal contract reduced the same downstream target by about 3.7 MiB versus the old fallback/detection configuration. NCCL remains required because CubeCL CUDA communication currently uses communicator, all-reduce, send, and receive APIs unconditionally.

Validation

• cargo fmt --all --check
• standalone cudarc feature contract tests
• CUDARC_CUDA_VERSION=12080 cargo check -p t4a-cubecl-cuda
• tenferro CUDA consumer check against revision 072cf82387ae099ce1d23efe6c0cbc68e231d9d2
• fresh controlled downstream release/debug measurements