Preserve small angles in batch rotation logarithmic map

[Doribelove]

axis_angles_from_matrices currently computes the angle with arccos((trace - 1) / 2). For a rotation of 1e-8 radians, the trace rounds to 3 and the batch logarithmic map returns a zero angle and the identity axis. This loses the rotation in batch processing.

Compute the angle with atan2 using both the trace and the norm of the antisymmetric matrix part, which remains informative near zero. This addresses the batch rotation portion of #377 without changing the scalar rotation or transformation code covered by #374 and #376. The API and precomputed traces/out options remain unchanged.

The regression tests cover zero, small positive and negative angles across seeded random axes, precomputed traces, an output buffer, a single matrix, and rotations near pi. The existing near-pi test now checks the rotation matrix and angle directly: exact pi rotations can legitimately use either axis sign, and the scalar arccos path can round a near-pi angle differently.

Validation:

• 98 passed in batch_rotations/test/test_batch_rotations.py and trajectories/test.
• Ruff check and git diff --check passed.
• Full suite in this environment: 3086 passed, 6 skipped, 7 failed. The same seven failures occur on an unmodified checkout of upstream develop (3084 passed, 6 skipped, 7 failed): one very tight 1e-16 rotation tolerance assertion and six tests requiring optional mesh dependencies.

Code and tests were developed with assistance from OpenAI Codex. The commit is signed off by 李永祺 doribelove@gmail.com.

[AlexanderFabisch]

Thanks for the contribution @Doribelove !