llama.cpp issue #28676: silent output corruption with HIP multi-GPU on Windows 11

gfx1100 + gfx1201 pair on Win11 with ROCm 10.0: fluent-but-nonsense word salad at healthy speed (llama-bench reports pp ~197, tg ~26), so benchmark numbers do not reveal it. Controls are clean: single-GPU HIP, dual-GPU Linux HIP, and Windows Vulkan. A second user (andj1210) confirms on RX 9070 / R9700 AI Pro with ROCm 10 and reports that -DGGML_CUDA_NO_PEER_COPY=ON makes it disappear. Tracked as a ROCm/AMD Windows driver issue rather than a llama.cpp bug.