Self-contained Colab notebook for the digit tutorial (POC) - #4

Proof-of-concept for lightweight, self-contained tutorial notebooks that run from the tutorial webpage and Google Colab — starting with the digit domain-adaptation example. One example now; the pattern rolls out to the rest one notebook at a time.

What this does

• Self-contained notebook. Reworks others/image_classification/digit_domain_adaptation/tutorial.ipynb to inline its config and model build and drop every sibling import (config.py, model.py, configs/tutorial.yaml). Colab loads only the .ipynb, so it previously could not run; now it can.
• Clean Colab setup. Replaces the old git clone … cd examples/digits_dann cell with pip install "pykale[image]" matplotlib (a no-op off Colab).
• Friendly + lightweight. Plain-language sections (what domain adaptation is, source vs. target, what DANN does); ~3 epochs on a fifth of the data → a few minutes on a free CPU. Points at the full main.py for real numbers.
• Site wiring. tutorial_actions now points Colab / Download-Notebook at each example's own tutorial.ipynb, and conf.py enables those buttons per-tutorial when the file exists. Examples without a notebook keep a disabled button with a reason — so this is additive and the rollout is incremental.

Publish target

Each example's notebook lives in its own folder (<example>/tutorial.ipynb), beside the narrative and example code.

Verification

• Executed the committed notebook headless end-to-end: no errors, USPS (target) accuracy ≈ 0.73, MNIST ≈ 0.94. Committed with no stale outputs.
• 96 tests pass; validate and the strict -W --keep-going build are green.

Notes

• The Colab button resolves once this is on main (URL is blob/main/…).
• The gallery narrative (plot_*.py) is unchanged; whether to converge on a single source (narrative ⇄ notebook) to avoid drift is a follow-up decision before the full rollout.

Draft for review.