Llama.cpp gets ANE backend for faster M-series inference

Unlocking the ANE is a watershed moment for the Mac AI ecosystem, finally providing a way to target the "black box" NPU that has historically been difficult to utilize outside of official CoreML workflows.

• –Uses private APIs to bypass the common CoreML "fallback to GPU" issue, ensuring matrix operations actually hit the ANE hardware.
• –Employs a tiered strategy where the ANE handles the computationally heavy prefill stage (N >= 64) while Metal or CPU takes over for token generation.
• –Performance on M4 Pro hits 4.0 TFLOPS peak, offering a significant efficiency boost and freeing up GPU cycles for other tasks.
• –The project is now on the official llama.cpp roadmap as a research priority, signaling potential upstream integration.
• –Clarifies that the optimization targets the standard ANE found in all Apple Silicon (M1-M4), not just the new "Neural Accelerator" cores exclusive to M5.