CI: simplify onboard scene workflows - #1682

Summary

• remove the legacy path-based a2a3 SDMA mode from the reusable NPU scene-test workflow
• make both ci.yml and ci-self-cpu.yml use the marker-based a2a3 SDMA split
• keep the onboard NPU DFX smokes in the existing st-onboard-* jobs instead of splitting them into extra jobs
• remove the redundant graphviz warning step from the a2a3 onboard workflow
• refresh the CI docs to match the latest main topology while documenting the onboard DFX behavior

Relation to #1639

#1639's main dedupe is already present on main: ci.yml and ci-self-cpu.yml share the canonical _detect-changes.yml, common job bodies are reusable workflow_call workflows, and the pip cache / venv scaffolding lives in composite actions.

This PR keeps that direction but limits itself to the remaining a2a3 onboard cleanup: remove the legacy SDMA path mode, keep the marker-based split, and document that the onboard DFX smokes stay inside the base onboard jobs rather than becoming separate CI jobs.

Checkout intentionally remains inline as actions/checkout@v5 inside the reusable workflows. A local composite action cannot safely wrap the first checkout step because local actions are loaded from the workspace after checkout has already happened.

Validation

• rebased onto the latest main (c9ccaf65 on August 6, 2026)
• parsed touched GitHub workflow YAML files with PyYAML
• ran git diff --check