Summary

Adds three capabilities needed for single-image inference with the SAM 3.1 multiplex checkpoint:

1. build_sam3_multiplex_tracking() — new public builder that assembles Sam3MultiplexTrackingWithInteractivity without the video-predictor wrapper. Exposes .detector (Sam3MultiplexDetector) for use with Sam3Processor-based image inference. Includes FA3 auto-detection (imports flash_attn_interface, falls back to use_fa3=False if absent).

2. Sam3VideoPredictor multiplex routing — when checkpoint_path contains "multiplex", __init__ now calls build_sam3_multiplex_tracking instead of build_sam3_video_model, which was incompatible with the multiplex checkpoint architecture (different tensor shapes in maskmem_backbone). Adds enable_inst_interactivity parameter (instance interactivity is always enabled in the multiplex architecture).

3. decoder.py forward_ffn dtype fix — casts tgt to the weight dtype before the autocast-disabled region, preventing dtype mismatches when torch.compile emits bf16 kernels (observed on CUDA 13 / Blackwell).

4. sam3/__init__.py — exports Sam3VideoPredictor and Sam3Processor from the top-level package.

Motivation

Sam3VideoPredictor loaded with sam3.1_multiplex_fp16.safetensors crashed with size mismatches because the multiplex checkpoint uses a different maskmem_backbone architecture than the non-multiplex model built by build_sam3_video_model. build_sam3_multiplex_tracking builds the correct architecture and loads the full checkpoint (tracker + detector) in one pass.

Test plan

• Load sam3.1_multiplex_fp16.safetensors via Sam3VideoPredictor — previously crashed with size mismatch, now loads the multiplex architecture correctly
• Confirm Sam3VideoPredictor with a SAM 3 (non-multiplex) checkpoint still routes to build_sam3_video_model unchanged
• Confirm build_sam3_multiplex_tracking() returns a model with .detector accessible to Sam3Processor
• Confirm bf16 inference on CUDA 13 no longer raises dtype errors in forward_ffn