Skip CUDA-only libhipcxx atomic_codegen when CUDAToolkit is absent

[Dillflix]

Motivation

• The TheRock configure step aborts on AMD-only systems because math-libs/libhipcxx/test/atomic_codegen unconditionally requires CUDA tools like cuobjdump even when CUDAToolkit/nvcc are not present.
• This is an NVIDIA-only test that should not block ROCm-only builds.

Description

• Added a YAML-driven patch in strix-halo/vllm-packages.yaml to insert an early-return guard into math-libs/libhipcxx/test/atomic_codegen/CMakeLists.txt when CUDAToolkit_FOUND is false so the CUDA-only test setup is skipped.
• Documented the issue and the workaround in strix-halo/BUILD-FIXES.md with symptom, root cause, and fix rationale.
• No other build logic was changed; the patch is applied via the existing manifest-driven patching mechanism.

Testing

• Ran git diff --check to ensure no whitespace or obvious diff problems and it passed.
• Inspected the updated strix-halo/vllm-packages.yaml and strix-halo/BUILD-FIXES.md with line-numbered output to verify the exact inserted patch and documentation.
• Attempted to validate the YAML patch entry with yq, but yq was not installed in this environment (informational only); the textual inspections above confirmed the changes.

---

Codex Task