Replace dict-literal result cells in the tutorials

Problem

25 code cells across the tutorials end in a hand-assembled dict literal used to print labelled
numbers. Examples:

{"flagged": sorted(flagged), "injected": sorted(truth["bad_images"])}
{"correlation with plate row, before": round(float(before), 3), "after": round(float(after), 3)}
{"constant features kept": [...], "duplicate pairs left intact": [...]}

Counts per notebook: 00 3, 02 1, 03 1, 04 1, 05 2, 06 3, 07 4, 08 4, 09 5, 10 1.

Reasoning

Two problems.

The output is not reusable. A reader cannot sort it, filter it, plot it or paste it into their own
analysis. Every one of these numbers came from a DataFrame or an array that was more useful before it
was flattened into a dict of prose keys.

The pattern is also the main reason the pages read as generated rather than written. No tutorial in
scanpy, squidpy, decoupler or pertpy prints results this way.

Most of these cells are also assertions rather than demonstrations: they compare a result against
uns["mantispy"]["truth"] to show the function recovered an injected effect. That is a unit test.
It belongs in tests/, and several equivalents are already there.

Proposed fix

For each of the 25 cells:

• If it demonstrates a result, end the cell in a DataFrame, a figure or an object repr
• If it asserts correctness against injected ground truth, move the check to tests/ and replace
  the cell with the figure or table a reader would actually look at
• Keep at most two ground-truth comparisons in the whole documentation, on Examples pages where
  "here is the answer, verified" is the point of the page

Depends on #44 and #45. Take one
notebook per agent.

Review dependencies #44 and #45, then choose one of the listed tutorial notebooks and inspect its dict-literal cells alongside equivalent checks in tests/. Replace demonstration cells with a DataFrame, figure, or object repr, move ground-truth assertions to tests/, and keep at most two documented comparisons overall.