fix(hip): map Gemma4 BSA runtime calls

[weicj]

Summary

This PR fixes a HIP build failure in the Gemma4 BSA prefill path.

After the Gemma4 backend landed, HIP builds can reach gemma4_graph.cpp and fail because that path uses CUDA runtime names directly, including cudaMemcpy2D and cudaDeviceSynchronize. Most of the project already routes these CUDA-style names through gpu_runtime_compat.h, but Gemma4 did not include that compatibility header and the header did not map the synchronous cudaMemcpy2D call yet.

Changes

• Add cudaMemcpy2D -> hipMemcpy2D to dflash/src/common/gpu_runtime_compat.h.
• Include common/gpu_runtime_compat.h from dflash/src/gemma4/gemma4_graph.cpp so the Gemma4 BSA path uses the same CUDA/HIP runtime compatibility layer as the rest of the C++ backend code.

Notes

• Verified on ROCm 6.3.3 / gfx906 with an initialized submodule checkout: HIP dflash_server builds successfully, and a Gemma4 E4B Q4 smoke run loads on hip:0 and returns a /v1/chat/completions response.
• This removes the Gemma4 HIP build blocker hit while validating the rebased mixed-backend work.

[cubic-dev-ai]

No issues found across 2 files

_{Re-trigger cubic}

[howard0su]

Please add HiP build into CI workflow to avoid such breakage.