DFlash in llama.cpp tested: 4.44x speedup on Qwen 3.6 27B

A user tested the newly merged DFlash speculative decoding in llama.cpp on Qwen 3.6 27B, achieving 4.44x faster inference at 36K context on an RTX 6000 PRO. DFlash uses a block diffusion drafter for acceleration.