fix: preserve small rotations in quaternion logarithms

[dddd-jh]

UnitQuaternion.Rx(1e-9).log().v currently returns [0, 0, 0] instead of [5e-10, 0, 0]. acos(s / norm) loses the angle when the quotient rounds to one; it also loses precision for larger small rotations. Both singleton and sequence logarithms are affected, including non-unit quaternions.

Use atan2(norm(v), s) in both paths. This is the same principal angle for a nonzero vector part, retains small-angle information, and preserves the branch for negative scalar parts. The existing zero-vector handling and scalar log(norm(q)) are unchanged.

Regression tests cover signed small rotations from 1e-12 to 1e-4 radians, an ordinary rotation, singleton/sequence agreement, non-unit scales, exp(log(q)), input preservation, and the branch near the negative real axis. The original implementation fails eight scalar subcases, both sequence subcases, and the negative-real precision check.

Validation on Windows / Python 3.12.14:

• Complete suite: 350 passed, 3 skipped, with all 12 new subtests passing, on both NumPy 2.5.3 / SciPy 1.18.1 and NumPy 1.26.4 / SciPy 1.15.3 (MPLBACKEND=Agg, the CI timeout options).
• Independently checked 100 deterministic rotation vectors against scipy.spatial.transform.Rotation, through singleton unit, singleton scaled, and sequence unit quaternion logarithms: all 300 comparisons pass after the fix; 270 fail a 1e-13 relative-error check before it. Maximum relative error after the fix is about 3.6e-16 in both dependency environments.
• Black 23.10.0, flake8's syntax/undefined-name checks (E9,F63,F7,F82), and git diff --check pass.
• Isolated sdist and wheel builds pass.

This validates the quaternion math and repository tests; no ROS nodes or robot hardware were run.