Pivot to llama.cpp with DFlash speculative decoding on ROCm

- Abandon vLLM/FP8 (Triton compilation hangs, ~0.5 tok/s on Strix Halo)
- llama.cpp v10133 with ROCm0 offload, DFlash draft model
- Main: unsloth/Qwen3.6-27B-MTP-GGUF:Q8_0 (~22GB)
- Draft: Alittlehammmer/Qwen3.6-27B-DFlash-GGUF-llama.cpp:Q8_0 (~1.8GB)
- 3 parallel streams, 131072 context/stream
- systemd service with minimal sandboxing (ROCm requires /dev/kfd access)
- Baseline: 7.62 tok/s (threshold 5.0 tok/s)

scripts/llama-server.service

@@ -0,0 +1,14 @@
+[Unit]
+Description=llama.cpp DFlash+MTP Inference Server (Qwen3.6-27B on ROCm)
+After=network.target
+
+[Service]
+Type=simple
+User=dank
+Group=dank
+ExecStart=/home/dank/llama.cpp/build/bin/llama-server --hf-repo unsloth/Qwen3.6-27B-MTP-GGUF:Q8_0 --spec-draft-hf Alittlehammmer/Qwen3.6-27B-DFlash-GGUF-llama.cpp:Q8_0 --spec-draft-n-max 16 --device ROCm0 --spec-draft-device ROCm0 --gpu-layers all --fit on --fit-target 512 --parallel 3 --ctx-size 131072 --batch-size 4096 --port 8080 --host 0.0.0.0 --threads 16 --load-mode mlock
+Restart=on-failure
+RestartSec=10
+
+[Install]
+WantedBy=multi-user.target