llama.cpp b10040 restores integrated GPU support on HIP builds

The llama.cpp project released build b10040, which includes a fix for HIP/ROCm builds to correctly handle the `prop.integrated` device property. Previously, a workaround for corrupted output on Jetson Orin devices unconditionally set this flag to false for all CUDA/HIP devices, causing an inconsistency that broke integrated-GPU host-buffer usage on AMD APU/UMA parts.

• The fix guards the Jetson workaround so it only applies to non-HIP (CUDA) builds.
• For HIP builds, the code now restores `prop.integrated` from `hipDeviceProp_t`, allowing `supports_buft()` to correctly accept CUDA host buffers on AMD integrated graphics.
• This resolves issue #23977 regarding broken integrated-GPU support on ROCm.

The release provides binaries for macOS (Apple Silicon and Intel), Linux (CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL), Android, Windows (CPU, OpenCL, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP), and openEuler.