Add MoE DFlash2 A/B: loses harder than dense (0.89x code), contradicting the prediction

Add DFlash2 speculative decoding A/B on dense 27B (net loss despite 76% acceptance)

Add oMLX Qwen3.8-27B dense 4bit at 8k ctx (2 sessions, GPU idle both times)

Add M2 Max 32GB benchmarks: MoE vs dense, MTP on Metal, KV quant, MLX vs llama.cpp