Add X3D video backbone with temporal input support

[dandarm]

Summary

• add lightweight X3D-based backbone that preserves the temporal dimension and decodes to existing heads
• emit both fused and spatio-temporal tensors from datasets/inference to drive 3D backbones
• document configurable backbone/temporal options for training
• update the train.sh launcher to show the new X3D arguments while keeping the legacy 2D ones as commented references

Testing

• not run (not requested)

---

Codex Task