llama.cpp b9930 fixes sequence position bug across all platforms

The llama.cpp project, a popular C/C++ implementation for running large language models locally, released version b9930 on July 8, 2024. This patch addresses a specific bug in `llama-batch` that allowed decreasing positions within a sequence (PR #25449), which could cause incorrect token ordering during inference. The fix ensures that position arrays maintain monotonicity within each sequence, a requirement for correct transformer model execution.

This release is notable for its breadth of platform support. Builds are available for macOS (Apple Silicon including KleidiAI-optimized, Intel x64), Linux (x64/arm64 with CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL), Windows (x64/arm64 with CPU, CUDA 12.4 and 13.3, Vulkan, OpenCL Adreno, HIP), and Android arm64. The commit is signed with GitHub's verified signature (GPG key B5690EEEBB952194). For users running local LLMs—from LLaMA, Mistral, to custom models—this patch prevents silent corruption in batch generation workflows.