test(notebooks): run the demo notebook end to end on both backends

[ggprior]

TL;DR: TODO

Claude generated description

The demo notebook had no execution coverage — only the nightly check that its
pip install lines resolve to the latest release. It had drifted: the
synthetic-data cell raised on the client backend, and three separate bugs sat
behind that one symptom, in three repos.

CI

tests/test_notebook_execution.py runs all 86 cells in a real Jupyter kernel,
once per backend, and checks the ROC AUC and RMSE the notebook prints still
clear a floor. Nightly, on dispatch, and on PRs touching the notebook. Failures
on the nightly open or comment on a tracking issue, matching the existing
install check; the executed notebook is uploaded as an artifact either way.

The notebook installs its own dependencies, so the workflow seeds the venv with
an editable install of the checkout first — uv pip install tabpfn audits an
already-satisfied requirement instead of pulling PyPI's wheel over the top,
which is what makes the run test the commit rather than the release.

Runtimes on a GPU runner: 4m30s local, 9m client.

Notebook fixes

Three things blocked an unattended run:

• The backend prompt. TABPFN_DEMO_BACKEND=local|client now skips it;
  anything else still prompts, so interactive use is unchanged.
• The token prompt. The local branch checked the Colab secret and then went
  straight to getpass, ignoring a TABPFN_TOKEN already in the environment.
  Environment first, then Colab secret, then prompt.
• gdown. The text-data cell calls it with --fuzzy, a flag current
  versions dropped, and nothing installs it — it only worked because Colab
  ships an older one preinstalled. Now installed, and called with -O so the
  output filename does not depend on Drive metadata.

Depends on

The client-backend job stays red until two upstream fixes ship, both surfaced by
this workflow:

• PriorLabs/tabpfn-extensions#390 — outlier detection calls .clone() on
  logits, which is a tensor locally and an array from the client.
• PriorLabs/tabpfn-client#369 — output_type="full" returns NaN where a
  local prediction has -inf, and raises above the full-output row cap instead
  of splitting.

Merge those, release both, then this.

Also checked

No notebook path needs a Hugging Face token. All five Prior-Labs/* model repos
are ungated and resolve anonymously, and a fresh HF_HOME with HF_TOKEN unset
downloads and predicts on TABPFN_TOKEN alone. The notebook only mentions
Hugging Face in prose.