llama.cpp b9257: Vulkan IM2COL shader optimization boosts GPU inference

The latest release of llama.cpp, tagged b9257, introduces a significant optimization to the Vulkan IM2COL shader. The IM2COL operation is a critical step in convolutional neural network inference—it rearranges image patches into matrix columns for efficient matrix multiplication on GPUs. By optimizing this shader, llama.cpp reduces latency and improves throughput for running large language models locally on Vulkan-compatible hardware (e.g., AMD, Intel, and some NVIDIA GPUs under Vulkan). The changes also include added comments and improved code formatting for maintainability.

The release provides pre-built binaries for a wide range of platforms: macOS (Apple Silicon and Intel, with optional KleidiAI acceleration), Linux (x86, ARM, s390x with CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16), Windows (CPU, ARM, CUDA 12/13, Vulkan, SYCL, HIP), Android (ARM64 CPU), and openEuler (x86 and aarch64 with ACL Graph). This broad support ensures developers and end-users on diverse systems can benefit from the IM2COL shader performance gains without recompiling. The optimization is particularly valuable for running local AI models on consumer-grade GPUs, making llama.cpp a stronger choice for privacy-preserving, on-device inference.