llama.cpp PR caches hot MoE experts on GPU, boosting speed

A llama.cpp PR (#26563) caches frequently used MoE experts in VRAM while cold experts run on CPU, reporting 33→56 tok/s with 8GB VRAM. A user offloading hot experts to VRAM saw 50% boost (20→30 t/s) on Qwen 3.8 Flash Next.