Multi-hour llama.cpp optimization experiments on Qwen MoE models, patches, benchmarks, and reproduction guides

I had Astra run a multi-hour investigation into speeding up local MoE inference while keeping the model weights and quantization unchanged. The experiments covered Qwen3.6-35B-A3B and Qwen3.8 Flash-Next. Some workloads showed substantial gains, particularly prompt processing and source-code editing.