llama.cpp b9686 fixes segfault on long prompts in eagle3

llama.cpp's latest release (b9686) addresses a critical segfault error in the eagle3 speculative decoding module that triggered crashes when processing long prompts. The spec: fix segfault error on long prompts for eagle3 (#24707) resolves a memory access issue that could halt inference unexpectedly, improving reliability for users running local AI models with speculative decoding. This technique accelerates generation by predicting multiple tokens in parallel, but the bug made it unstable with lengthy inputs.

Supported platforms include macOS (Apple Silicon and Intel), iOS, Linux (x64, arm64, s390x) with backends like Vulkan, ROCm 7.2, OpenVINO, SYCL, and CUDA (12/13). Windows builds cover CPU, CUDA, Vulkan, OpenVINO, SYCL, and HIP. Android arm64 is also included. The release is signed with GitHub's verified signature, ensuring authenticity. This update is crucial for developers and hobbyists running local LLMs, as speculative decoding is key to reducing latency in chat applications and code assistants.