fix(space_ops): flip the axis when angle_axis_from_quaternion folds the angle

[dylanpulver]

Overview: What does this pull request change?

angle_axis_from_quaternion (manim/utils/space_ops.py:143) folds an angle above PI with

if angle > TAU / 2:
    angle = TAU - angle

but leaves the axis untouched. A rotation by angle about axis is a rotation by TAU - angle
about -axis, so for every quaternion whose angle exceeds PI the returned pair describes the
inverse rotation. This adds axis = -axis to the fold, plus a test.

Motivation and Explanation: Why and how do your changes improve the library?

On main, feeding quaternion_from_angle_axis's own output back in does not round trip:

axis = np.array([0.0, 0.0, 1.0])
q = quaternion_from_angle_axis(4.0, axis)
angle, ax = angle_axis_from_quaternion(q)      # -> 2.2832, [0. 0. 1.]
np.allclose(rotation_matrix(4.0, axis), rotation_matrix(angle, ax))   # False

5000 random (angle, axis) pairs, uniform angle over [0, TAU): 2556 fail the round trip on
main, 0 after. The split is clean — all 2556 have angle > PI, none have angle <= PI. The
neighbouring rotation_matrix_from_quaternion and rotation_matrix_transpose_from_quaternion were
run over the same 5000 inputs as a control: 0 failures each, before and after.

Which of the two possible remedies to take was settled by scipy, already a hard dependency and what
rotation_matrix in this same module builds on: Rotation.as_rotvec returns exactly the
angle in [0, PI] + signed-axis form, and the patched function agrees with it on all 5000 cases.
Simply deleting the fold also repairs the round trip, but changes the returned angle range — happy
to go that way instead if you prefer it. No test caught this because the function has zero call
sites inside manim/ and had no test.

Matched pair, same command and env both times (uv run pytest tests/module -q -p no:randomly):
492 passed / 90 failed before → 499 passed / 90 failed after. The 90 are pre-existing here
because this machine has no LaTeX (FileNotFoundError: 'latex'); they are identical on an
unmodified checkout. pre-commit run --files ... passes all hooks at the pinned versions
(ruff 0.16.2 lint + format, mypy, codespell).

Two mutants, both with the source marker asserted before measuring:

• revert the source, keep the test → the 4 cases above PI fail, the 3 at or below PI pass;
• drop the fold entirely instead → the rotation-matrix assertion passes and the [0, PI]
  assertion fails, which is what pins the remedy rather than just the round trip.

Links to added or changed documentation pages

angle_axis_from_quaternion in the space_ops reference; its Returns text now names the angle
range.

Further Information and Comments

Not tested: the LaTeX-dependent tests did not run here. I have not touched the -> Sequence[float]
return annotation, which does not describe the (float, ndarray) tuple, to keep the diff minimal.

Written with AI assistance (Claude); the measurements above were run locally on this branch.

Reviewer Checklist

• The PR title is descriptive enough for the changelog, and the PR is labeled correctly
• If applicable: newly added non-private functions and classes have a docstring including a short summary and a PARAMETERS section
• If applicable: newly added functions and classes are tested