Fix gaze estimation device timestamp propagation

[PetrNovota]

Summary

• propagate device timestamps through gaze-estimation crop configuration groups
• preserve device timestamps on derived crop configs, head-pose tensors, and annotations
• document the timestamp requirement for future changes

Problem

The three FrameCropper synchronization nodes received config groups with a default device timestamp of 0 ms while camera frames carried valid device timestamps. This prevented synchronization and caused continuous Sync warnings.

Testing

• reproduced the zero-timestamp Sync warnings on an OAK4 before the fix
• ran oakctl app run neural-networks/face-detection/gaze-estimation/ on an OAK4 after the fix; no Sync warnings appeared during the observation window
• Python compileall
• AGENTS.md validation
• generated index check
• Ruff, Ruff format, and mdformat pre-commit hooks