llama.cpp b9788 adds SYCL tensor parallelism for dual-GPU setups

The llama.cpp release b9788 introduces support for tensor parallelism via the --split-mode tensor flag in the SYCL backend. This implementation enables dual-GPU communication by adding comm_init, comm_free, and comm_allreduce_tensor functions to the meta-backend. For two devices, it uses a ring all-reduce strategy that switches between FP32 direct memcpy for small tensors and BF16 compression for larger ones. The code avoids OneCCL due to its single-device-per-process limitation, instead using persistent buffers to maintain SYCL pool invariants. Performance tests on dual Intel Arc Pro B70 GPUs show significant speedups over layer mode for Llama-3.3-70B and Qwen3-Coder-Next-80B-A3B models. The update includes new binaries for macOS, Linux, Windows, Android, and openEuler across CPU, CUDA, ROCm, Vulkan, and SYCL targets.