llama.cpp b9965 improves ARGSORT performance for small tensors

The llama.cpp project released version b9965, which includes a significant optimization for the ARGSORT operation on small tensors via the hexagon backend. This update introduces an efficient bitomic sort implementation using HVX registers capable of handling up to 1024 elements.

• Added efficient bitomic sort in HVX registers for up to 1024 elements.
• Fixed inverted VRORs in the hex-sort module.
• Specialized sort functions for common cases and added tracing/local context.
• macOS Apple Silicon (arm64) builds with KleidiAI are now disabled.

This release provides updated binaries for macOS, Linux, Android, Windows, and openEuler across various CPU and GPU backends including CUDA, Vulkan, ROCm, and OpenVINO.