docs: add runnable tiny training smoke tutorial

[2515830671]

Related Issue

Closes #41

Summary

Turn the existing tiny CLI smoke command into a comprehensive step-by-step
tutorial. The smoke test command was already documented briefly in
quickstart.rst and training.rst, but there was no standalone guide
covering prerequisites, expected output, artifacts, TensorBoard inspection,
and common failures in one place.

Changes

• New file: docs/cookbook/tiny-training-smoke.rst — a complete tutorial covering:
  • Prerequisites (GPU/CUDA requirement, CPU skip language, areno check, repository checkout)
  • The smoke test command with a per-flag explanation table
  • Expected output (training config, rollout logs, train logs with train_stats, step timing)
  • Expected artifacts (checkpoints with --save-path, TensorBoard metrics with --metrics-log-dir)
  • Common failures table (CUDA unavailable, missing areno_accel, flash-attn on T4, OOM, wrong working directory, dataset download issues, ratio_mean=1.0 / grad_norm=0.0 explanation)
  • Explicit "what the smoke test does NOT verify" section
  • Summary checklist table
• Modified file: docs/index.rst — added cookbook/tiny-training-smoke as the first entry in the Cookbook toctree

Acceptance Criteria (from issue #41)

Criterion	Status
Tutorial uses an existing example dataset/reward path	Uses gsm8k:main, examples/math/dataset_loader.py, examples/math/math_verify_reward.py
GPU requirement and CPU skip language are explicit	Hardware section + warning admonition
Prerequisites, expected logs, artifacts, common failures	All four sections present
Explains that it verifies wiring, not model quality	Top note + dedicated section
TensorBoard inspection when --metrics-log-dir is set	Full section with namespace/chart table
No code changes	Only .rst documentation files

Type of Change

• Documentation update
• Bug fix
• New feature
• Breaking change

[2515830671]

#41