Summary:
SAM 3.1 multiplex video tracking for the OSS sam3 repository.

Features:
- Multiplex architecture: batch N objects into fixed-capacity buckets for efficient multi-object tracking (16 objects per bucket)
- Unified API: build_sam3_predictor(version="sam3.1") with auto-download from HuggingFace (facebook/sam3.1)
- Flash Attention 3, torch.compile support (~31 FPS compiled on H100)
- Text prompts, point prompts, box prompts, object removal, refinement
- Clean inheritance: Sam3MultiplexBase extends Sam3VideoBase
- Shared base predictor (Sam3BasePredictor) for both SAM3 and SAM3.1

Performance (H100, 5 objects, compiled):
- SAM 3.1: ~31 FPS
- Memory: ~17.7 GB

Usage:
  from sam3 import build_sam3_predictor predictor = build_sam3_predictor()  # auto-downloads SAM 3.1 from HF

Pull Request resolved: https://github.com/facebookresearch/sam3/pull/503

Reviewed By: josephgreer

Differential Revision: D98500763

Pulled By: arpitkalla

fbshipit-source-id: e2b74568f24bb4ecfb3144a4e8ec559aae6db5e8