llama.cpp b9922 adds n_keep_tail for recurrent models

The llama.cpp project has released version b9922, which introduces a new `n_keep_tail` parameter in the `split_equal` function specifically for recurrent models.

• macOS Apple Silicon (arm64) binaries are available.
• macOS Intel (x64) binaries are provided.
• iOS XCFramework is included.
• Ubuntu x64 and arm64 CPU builds are released.
• Ubuntu Vulkan, ROCm 7.2, OpenVINO, SYCL FP32, and SYCL FP16 variants are available.
• Android arm64 CPU binaries are provided.
• Windows x64 and arm64 CPU builds are released.
• Windows CUDA 12.4, CUDA 13.3, Vulkan, OpenVINO, SYCL, and HIP builds are available.
• openEuler x86 and aarch64 builds with ACL Graph support are included.
• A standalone UI package is provided.

This release updates the binary distribution for llama.cpp across multiple platforms and hardware accelerators.