Add confidence-aware SO3 geodesic likelihood updates

[FlorianPfaff]

Summary

• add log-likelihood updates for SO(3)^K product particle filters
• add masked, confidence-aware, heteroskedastic geodesic likelihood helpers
• wire the partitioned SO(3)^K filter through component/block log-likelihood updates
• add regression tests for zero-likelihood log updates, confidence-to-noise mapping, per-component noise, and partitioned confidence-aware updates

Notes

This upstreams the generic detector-style likelihood machinery from SO3-Pose-Filter without porting SMPL/AMASS-specific code.

Testing

• Not run locally: the execution environment cannot clone GitHub repositories because DNS resolution for github.com is unavailable.