Feature: Implement Fine-Tuned SAM 3 Spleen Models

[blayyyyyk]

Description

As an ML engineer, I want to load our team's custom fine-tuned SAM 3 spleen models into the browser pipeline so that the segmentation accurately and exclusively targets the spleen.

Acceptance Criteria

• Convert the highest-performing fine-tuned spleen SAM 3 PyTorch model to the required web-compatible format (ONNX).
• Update the inference script to fetch and initialize these custom weights instead of the base SAM 3 model.
• Map NiiVue canvas click/drag events to the SAM 3 prompt inputs (transforming canvas coordinates to spatial model dimensions).
• Render the resulting binary mask returned by the model back onto the NiiVue volume as a colored overlay.

Implementation Notes & Resources

• Managing the coordinate space transformation between the NiiVue canvas and the SAM 3 image encoder is the critical path here.
• Use NiiVue's built-in overlay methods (e.g., nv.addVolume where the volume is designated as a label map) to display the returned mask.