DFlash in llama.cpp achieves 4.44x speedup on Qwen 3.6 27B at 36K context

DFlash speculative decoding merged into llama.cpp provides a 4.44x speedup over autoregressive decoding on an RTX 6000 PRO with 36K context length. Tested on Qwen 3.6 27B, the new block diffusion drafter outperforms previous MTP methods.