feat: npu workflow

[maciejmajek]

feat: NPU inference setup (FastFlowLM + AMD Ryzen AI)

Adds an optional, one-command setup for running LLM inference on the AMD Ryzen AI NPU via FastFlowLM, targeting the Ryzen AI MAX+ 395 (Strix Halo) machines.

What's in this PR

• pixi run -e local install-npu-inference — new optional task under the local-inference feature (pixi.toml).
• scripts/install_npu_inference.sh — idempotent installer that:
  1. NPU drivers — adds the lemonade-team/stable PPA and installs libxrt-npu2 + amdxdna-dkms; skips if already present, flags reboot when freshly installed.
  2. FastFlowLM (from source) — installs build deps (ninja, ffmpeg libs, libxrt-dev, etc.), clones RobotecAI/FastFlowLM into inference/FastFlowLM/, builds with the linux-default CMake preset, and sudo cmake --installs flm. Skips if flm is already on PATH.
  3. memlock limit — checks ulimit -l; appends * soft/hard memlock unlimited to /etc/security/limits.conf if not already unlimited, and flags reboot.
  4. Validate — runs xrt-smi examine + flm validate, capturing output. On success prints a clean ✓ NPU visible & ready / ✓ FLM ready; on failure prints the captured diagnostics plus a troubleshooting block (kernel module loaded? xrt-smi seeing the device?) with links to the FastFlowLM and Lemonade Linux guides.

Tested on

AMD Ryzen AI MAX+ 395, Ubuntu 24.04, kernel 6.17.0-1025-oem.

TODO

• Wire FastFlowLM into the actual AI stack (RAI) — add an flm serve task analogous to the serve-llm / serve-vlm llama.cpp tasks, exposing an OpenAI-compatible endpoint.
• Point the agents/orchestrator config at the NPU endpoint and run an end-to-end demo on NPU inference.
• Pick + document the model(s) served on the NPU (and download path).
• Decide whether to fold modprobe persistence + IOMMU check into the script or keep them as manual/documented steps.
• Document install-npu-inference in the setup docs (docs/setup_*.md).