Polish lab notebooks: hide scaffolding, grid_sample exercise, clearer figures

[roulbac]

Summary

• Mark prose/figure/checker cells hide_code=True across all lab notebooks so learners land on the exercise cells, not the scaffolding.
• Rework lab 0 exercise 1 (bilinear_sample) around a single F.grid_sample call — notebook text, hints, and labs/solutions/lab00.py — and add a CI test pinning the solution to grid_sample(align_corners=False, padding_mode="zeros") semantics and coordinate gradients.
• Clarify lab 0's interactive figures:
  • arrows are trimmed to stop at dot edges, so arrowheads are no longer buried under the sample dots (intro comparison, head demo, cross-attention diagram);
  • bilinear demo: the active unit cell is shaded, corner/point labels are placed outward and flipped inward near edges so nothing overflows the card at any slider position, and the table now reports the in-image weight sum (showing the fade to zero past the border) with a stable layout;
  • untrained-head demo: sample dots carry a small k index matching the fixed legend;
  • pyramid demo: per-level captions shortened so adjacent panels can no longer collide.

Test plan

• pytest tests/test_labs.py — 14 passed, 8 skipped (GPU-only).
• marimo export html labs/00_the_idea.py executes every cell headlessly with no errors; figures render.
• Label-geometry sweep over the full slider ranges (bilinear demo) and all 10 seeds (head demo): no text or marker leaves the canvas.