Support multiple box prompts in video inference

[oValach]

Summary

Sam3VideoInference.add_prompt() accepts box inputs with shape N x 4, but the current video inference path requires an initial visual prompt to contain exactly one box.

This change allows multiple initial box prompts to be packed into a single Prompt, while preserving the existing single-box and refinement behavior.

The multi-box representation uses the existing Prompt.append_boxes() path also used by the multiplex video implementation.

Motivation

Detector-driven video pipelines commonly produce multiple boxes for the same frame. With the current API, these boxes need to be submitted individually, which results in one single-frame inference call per box.

Supporting multiple boxes in one prompt allows all box prompts from a frame to be submitted together and processed in a single inference call.

Changes

• Support multiple initial box prompts in Sam3VideoInference.
• Preserve the existing single-box path unchanged.
• Preserve existing behavior when a visual prompt already exists.
• Add unit tests covering:
  • single-box visual prompts,
  • multiple boxes packed into one prompt,
  • existing visual-prompt/refinement behavior.

Validation

Tested with:

python test/test_video_inference.py -v
python test/test_video_predictor.py -v