MoE 54% speedup

This is the result of tuning Qwen3.6-35B-A3B GGUF UD-Q4_K_M (22.1GB) on an RTX 5070 Ti (16GB), Ryzen 9800X3D, 32GB RAM, llama.cpp b8829 (CUDA 13.1, Windows x64).

The key is using --n-cpu-moe 20 instead of --cpu-moe.

• --cpu-moe: offloads all MoE experts to CPU, using only 3.5GB of VRAM
• --n-cpu-moe 20: with 40 layers total, only the first 20 layers go to CPU while the rest are loaded onto GPU, utilizing up to 12.7GB of VRAM
• --n-cpu-moe 20 + -np 1 + 128K ctx: generation 79.3 t/s, prompt 135.8 t/s, VRAM 13.2GB

Based on 300-token generation, --n-cpu-moe 20 reached 78.7 t/s, a 54% improvement over --cpu-moe's 51.2 t/s, and prompt speed also rose from 87.9 t/s to 100.6 t/s.

The 128K context was added at essentially no cost by reducing recurrent-state memory with -np 1, and the example run includes -ngl 99, -fa on, -ctk q8_0, -ctv q8_0, --reasoning-budget -1, etc.