[Phase 2 follow-up] Produce retained visual baselines for pilot scenes

[alonshoa]

Split out from #4.

Objective

Produce retained baseline artifacts for the migrated Phase 2 pilot scenes so later manual visual regression review has concrete reference material.

Scope

• Render the three pilot scenes in the devcontainer using their cataloged render commands:
  • manim-slides render -ql decks/matrix_work/vecs_slides.py VectorsABtoV
  • manim -ql decks/losses/cross_entropy.py BinaryCrossEntropyLoss
  • manim -ql decks/matrix_work/parametric_curve_3d.py ParametricCurve3D
• Capture representative preview frames for each pilot:
  • Hebrew title/axes/vector addition/final loop for VectorsABtoV
  • Formula/positive-class curve/negative-class curve for BinaryCrossEntropyLoss
  • Initial axes/mid-curve/final helix for ParametricCurve3D
• Store curated baseline artifacts under the existing baselines/ directories.
• Keep full generated Manim output (media/, slides/) out of Git.

Acceptance criteria

• Each pilot baseline directory contains retained visual artifact files, not only README placeholders.
• README notes in each baseline directory match the captured artifacts.
• Baseline artifacts are small enough to review in Git and suitable for later manual visual regression comparison.
• A maintainer can reproduce the captures from documented commands inside the devcontainer.

Out of scope

• Adding automated visual regression gates to normal CI.
• Refactoring scene code.
• Importing additional legacy scenes.