llama.cpp b9936 fixes prompt batch splitting with min-step respect

The llama.cpp project, maintained by ggml-org, released version b9936 on July 9th. This is a minor but important bugfix release for the widely-used local AI inference engine. The core change addresses a server-side issue: when splitting prompt batches for processing, the server now properly respects the min-step parameter. Previously, batches were divided without considering min-step, which could lead to inefficient or incorrect prompt handling, especially for tasks requiring precise token generation steps.

This release also includes the usual comprehensive build matrix, supporting macOS (Apple Silicon and Intel), iOS, Linux (x64, arm64, s390x), Android arm64, and Windows (CPU, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP). Backends include ROCm 7.2 for AMD GPUs and KleidiAI for Apple Silicon. The fix ensures more reliable batch processing in production server deployments, particularly for users running local LLM inference with custom step configurations.