Compact verbose outputs in published notebooks

A full navigation audit found 24 retained DataTree representations, 20 progress streams, and 17 oversized or repeated model/table outputs in current published notebooks. These make otherwise valid pages slow and difficult to scan.

Highest-priority pages:
- variational_inference: one 141 KB progress stream, two more progress streams, and three tree dumps;
- getting_started: the first sample call renders an 83-line DataTree and large xarray HTML instead of a compact result;
- scientific_workflow_hssm and plotting: full tree dumps plus duplicated large dataframes or retained progress;
- choice_only and RLSSM learning pages: repeated progress, model, tree, and oversized summary outputs.

Acceptance criteria:
- disable retained sampling/training progress in static artifacts;
- assign sampling results and show targeted az.summary tables or concise group/shape status instead of full DataTree representations;
- suppress accidental Axes-array reprs;
- collapse repeated model specifications into one compact comparison where appropriate;
- migrate each legacy Jupyter notebook to marimo when its source is updated, in bounded page-family PRs;
- add focused output-hygiene checks without hiding legitimate diagnostics or plot payloads;
- regenerate full outputs, inspect figures, and validate strict docs for each migrated batch.

Implement in phases, starting with central first-user pages; do not combine all affected notebooks into one PR.

Start by inspecting the notebook sources for variational_inference and getting_started, then review the related scientific_workflow_hssm, plotting, choice_only, and RLSSM learning pages. Work in a bounded page-family phase, adding focused output-hygiene checks and migrating each updated legacy notebook to marimo. Regenerate full outputs, inspect figures, and validate strict docs when the batch is complete.

Written by the indexing model from the issue text.