feat: release 4.0 (npu workflow)

[maciejmajek]

Development → main

This PR replaces the manual multi-step setup with a pixi-managed workflow: environments per deployment profile and tasks covering the full lifecycle, from pixi run setup to pixi run demo. It also adds a config-driven inference dispatcher with NPU support, agentic conversation tracing, a custom spawn service in the sim, and CI.

Notable changes

• Pixi workflow: per-profile environments (sim, hmi, hil, single-pc gpu/npu) with tasks for the whole lifecycle: setup, build-sim, build-ros2, build-llama, build-fastflowlm, download-models, then sim, stack, inference, agents, hmi, plus demo, demo-trace, smoke-test, and dev tasks (lint, test, kill). ROS 2 deps come from pixi packages instead of submodules; O3DE switched to the SDK; versions locked in pixi.lock.
• Inference dispatcher (rai_app/inference/serve.py): resolves [endpoints.*] from config.toml into backend launch commands (llama.cpp for GPU/CPU, FastFlowLM for NPU), including health checks and weight downloads.
• NPU support: FastFlowLM submodule plus backend-aware structured output falling back to GBNF grammar where function calling is unavailable.
• Agentic trace: ConversationFileCallback records agent conversations to a run dir; demo-trace runs the full demo unattended and collects the trace, also dockerized (Dockerfile.demo).
• Sim: custom ROS 2 SpawnEntity service handler in the gem, prefab/asset updates, unused level removed.
• CI: workflow to build, publish, and smoke test Docker images.
• Docs: rewritten around the pixi flow.

Note:
28K additions come from pixi.lock file

[knicked]

I tested everything locally and in the docker pulled from the dockerhub.
The demo runs great with the NPU hardware and local LLMs.