llama.cpp b9555 fixes audio model im2col on Apple Silicon

The open-source llama.cpp project (116k stars on GitHub) dropped b9555, a maintenance release that resolves a specific Metal shader bug in the im2col (image-to-column) operation for 1D convolution. This bug primarily affected audio models — such as OpenAI's Whisper — that rely on 1D convolution layers, causing incorrect inference on Apple devices. The fix ensures correct computation on macOS and iOS Apple Silicon (arm64) when using the Metal backend. b9555 also ships pre-compiled binaries for a sprawling array of platforms: Windows (x64/arm64 with CUDA 12 & 13, Vulkan, HIP), Linux (x64/arm64 with Vulkan, ROCm, OpenVINO, SYCL), macOS (x64 and arm64, with optional KleidiAI acceleration disabled by default), Android arm64, and even IBM s390x. The release is signed with a verified GPG key and includes updated UI assets. For developers building audio AI applications on Apple hardware, this fix corrects a silent performance regression introduced in prior versions. The patch is minimal — focused solely on the Metal shader — but restores correct audio model inference without workarounds. Users running local Whisper transcription or custom audio models on Macs should upgrade immediately to avoid wrong outputs. The release also continues llama.cpp's trend of broad hardware support, making local inference practical across diverse consumer and enterprise GPUs.