Qwen3.6 MTP graft speeds llama.cpp

This is a strong proof-of-concept for turning a model's built-in MTP training into real local inference speed, not just another speculative-decoding experiment.

• –Qwen3 was trained with 3 MTP steps, so the draft heads are aligned with the base model instead of being a generic proxy drafter.
• –Keeping the 3 MTP layers at Q8_0 while the base stays low-bit is a sensible tradeoff: small memory cost, better draft acceptance.
• –The main limiter is ecosystem support, since it still depends on an unmerged llama.cpp PR rather than stock mainline builds.
• –If the reported acceptance rate holds across prompts, this is one of the more practical ways to make 27B-class local models feel much faster.
• –The release is also useful as a recipe: it includes the grafting script, raw layer extraction, and build steps for reproducing the setup on other models.