Strix Halo patch speeds MoE prefill

This is the kind of hardware-specific optimization that matters a lot to the people actually running local inference on Strix Halo, even if it never becomes a general-purpose llama.cpp feature.

• –The PR targets gfx1151 and reduces VGPR pressure by changing MMQ tile sizing and warp counts; that is why the win is concentrated on this AMD path.
• –The effect is strongest at low context because prefill dominates there; as context grows, flash attention takes over and the uplift shrinks.
• –The patch is MoE-specific, so dense models should not be expected to see the same uplift.
• –The numbers are good enough to justify a local fork if you own the hardware, but not broad enough to treat as a universal llama.cpp optimization.
• –For developers benchmarking local AI stacks, this is a reminder that “rejected upstream” does not mean “useless”; it can still be a worthwhile distro patch for a narrow machine/model combo.