fix(hip): restrict wavefront guard to device pass

[pepuscz]

Summary

Restrict the rocWMMA wavefront-size compile-time check to the HIP device pass.

Some ROCm 7.1 host passes expose __AMDGCN_WAVEFRONT_SIZE without making it an integer constant expression. Evaluating it in a preprocessor arithmetic condition then rejects an otherwise valid gfx1151 build. The device pass still performs the existing wave32 safety check, so wave64 targets continue to fail loudly.

Published qualification patch: lucebox-rocm71-host-wavefront.patch

Validation

• git diff --check
• ROCm 7.1.1 / gfx1151 qualified build completed with this guard
• The patch changes only preprocessor scope; kernel code and device-pass validation are unchanged

[cubic-dev-ai]

No issues found across 1 file

_{Re-trigger cubic}