Write a runnable tiny training smoke tutorial

[adohe]

Spec candidate: D2.

Scope

• Turn the tiny CLI smoke command into a tutorial with prerequisites, expected logs, expected checkpoint/metrics artifacts, and common failures.
• Explain that it verifies wiring, not model quality.
• Include TensorBoard inspection when --metrics-log-dir is set.

Acceptance

• Tutorial uses an existing example dataset/reward path.
• GPU requirement and CPU skip language are explicit.
• No code changes.