Add threshold input to filter low-confidence detections

[jacoboromerodiaz]

While using this model I noticed it sometimes detects objects that aren't actually present in the image. These false detections still pass the current hardcoded box_threshold=0.3 in detect(), meaning the threshold alone isn't always tight enough.

This PR exposes box_threshold as a user-facing threshold input in the Replicate predictor, so callers can tune it without touching the code. The default stays at 0.3 to preserve existing behavior.

Changes:

• Added threshold: float Input to predict.py (range 0.0–1.0, default 0.3)
• Threaded it through run_grounding_sam, both detect() calls (positive and negative prompts.