llama.cpp's b10359 adds WebGPU fixes, i32 support

ggml-org's llama.cpp, the go-to C++ library for running LLMs locally, released b10359 on August 11. This incremental update focuses on stabilizing the WebGPU backend by fixing CI errors referenced in issues #25025 and #25262. The commit also rebases and refines the new flash attention test, disabling subgroup matrices when max_kv_tile equals zero, and removes stray log output. A notable functional change is added i32 (32-bit integer) support to the cpy operation, which now enables the all-ops test to pass—a step toward broader tensor type compatibility for edge cases.

The release also restores non-target CI tests and comments out a TODO in build-cpu.yml, tightening the project's continuous integration pipeline. llama.cpp continues to support an extensive range of hardware and OS targets, with prebuilt binaries for macOS (Apple Silicon and Intel), iOS, Linux (x64, arm64, s390x, plus Vulkan, ROCm 7.14, OpenVINO, and SYCL variants), Android, Windows (CPU, CUDA 12/13, OpenCL Adreno, Vulkan, and more), and openEuler. With 124k GitHub stars, llama.cpp remains the most popular tool for on-device LLM inference, and b10359 reinforces its stability across diverse deployment scenarios.