Preserve small rotations in quaternion-to-axis-angle conversion

[sylvesterkaczmarek]

Summary

Fixes #563.

Compute the principal angle with atan2 using the quaternion's vector magnitude. Arccos of the scalar loses valid small rotations once that scalar rounds to one, returning zero even above the existing tolerance.

Retain the principal-angle range, negative-pi half-turn convention, subthreshold zero output and scalar-range validation. Input arrays are not modified.

Validation

Twenty-seven regression cases cover small and ordinary angles, equivalent quaternion signs, identity/subthreshold cases, half turns and invalid scalar controls. Expected rotations are checked analytically and against MuJoCo's actual mju_axisAngle2Quat/mju_quat2Vel functions, with rotation-matrix roundtrips.

• New tests on unchanged main: 9 failed, 18 passed.
• Selected baseline: 135 passed.
• dm_control/utils and dm_control/composer/variation: 162 passed on Python 3.11.16 and 3.12.11, with MuJoCo 3.14.0.
• Pylint errors-only passed with MuJoCo C-extension inspection enabled. Pyink checks on changed source lines/new tests, compilation and git diff --check passed.

Validation was local on macOS. One existing zero-axis-angle RuntimeWarning remains unchanged. Tests use native conversion routines, not rendering or the complete control-suite simulation matrix. Nonunit quaternion semantics and rotations below the existing tolerance are outside the intended change.

Based on main at a04e3e4cf56c12117d2294bb090f9acec21e5c67. No dependencies or workflows change.