Add temporal window stacking for early fusion inputs

[dandarm]

Summary

• introduce a configurable temporal_T parameter and shared selector to build centered temporal windows with fallbacks
• extend datasets, training, and inference to stack neighboring frames along the channel dimension for early fusion
• adapt the model stem to accept 3*T channels while keeping the rest of the architecture unchanged and backward-compatible
• add train_temporal.sh helper to launch temporal training with temporal_T

Testing

• Not run (not requested)

---

Codex Task