feat: allow point-prompt initialization by loading backbone

[aniketsrinivasan]

PR: Enable SAM3 image point-prompting in sam3_facebook

Summary

• Fixes missing wiring for SAM1-style interactive point prompting by ensuring the interactive tracker is constructed with a visual backbone when enable_inst_interactivity=True.

What changed

• model_builder.py
  • Build the interactive tracker with with_backbone=True (so SAM3InteractiveImagePredictor.set_image(*) can call forward_image and compute embeddings).
  • Extend checkpoint loading to also map detector.backbone.* weights into inst_interactive_predictor.model.backbone.* when instance interactivity is enabled.

Why

• In the upstream sam3 repo, point prompting works because the interactive predictor wraps a tracker that includes a backbone and receives appropriate weights. sam3_facebook previously built the tracker without a backbone, preventing feature initialization needed for point prompts.

Notes

• No behavior changes to non-interactive inference paths.