Qwen3.8-Flash-Next on 2x3090 + DDR4, part 4: 2.2-2.5x faster prefill by kicking the expert cache off the GPU while the prompt runs

Part 4 of the same box. Part 1 was 17 -> 25-29 t/s with the expert cache PR, part 2 was 37-41 t/s after switching to UD-Q4_K_XL and stacking MTP on the cache, part 3 was the top-k fallback that was sorting more than it needed to. This one is all about prefill, which was honestly the weak spot the wh