Backlog: CUDA follow-up work

[sunnycase]

Context

Follow-up backlog for the CUDA target/runtime work introduced by PR #1.

The current PR adds the CUDA target, native CUDA runtime module support, CUDA-aware NTT runtime primitives, CUDA kernel tests, and a dedicated Linux CUDA CI job. The items below are intentionally tracked outside the PR so they do not block the initial CUDA support merge.

Backlog

• Register and maintain a dedicated self-hosted CUDA runner pool so the CUDA job can run as a required PR check.
• Expand CUDA CI from the focused kernel coverage toward the full CUDA kernel test class once runner capacity and runtime stability are sufficient.
• Add matrix coverage for additional CUDA architectures or toolkit versions if compatibility across GPU generations becomes a release requirement.
• Continue reducing duplicated CPU/CUDA codegen and runtime paths as the NTT target abstraction stabilizes.

Notes

• Full UnitTestCUDAKernels execution requires a Linux runner with an NVIDIA GPU, CUDA toolkit, nvcc, clang/clang++, and the labels self-hosted, linux, x64, cuda.
• The initial CUDA CI job uses CUDA architecture 120, matching the local validation environment used for PR Add CUDA Target, Runtime, and Kernel CI Support #1.