Export artifacts workflow and decoder export fixes

[rbavery]

Summary

• end to end single pt2 export of SAM3 model, including
  • image encoder, text encoder, encoder feature map fusion, decoder
• add integration tests for each sub-module and comfirm output equality to eager mode.
• tests input dynamism for batch size, number of prompts, and variable image size is supported with an extra argument for a padding mask to pad to fixed size.
• per module and export/load timing benchmarks
• allow configuring num_feature_levels, which dictates how many feature maps the encoder produces. more feature maps = better performance on different object scales. This exposes this config export and benchmark script CLI flags. but only for some modules, this is not supported for the end to end SAM3 pipeline right now

Timing Results (CUDA 3090 24GB)

Feature levels = 1
Input image: 1008x1008 (fixed shape that sam3 supports by default)

• Inference (full pipeline): 0.435s

• Inference (image encoder): 0.297s

• Inference (text encoder): 0.011s

• Inference (encoder fusion): 0.081s

• Inference (decoder only): 0.050s

• Load full pipeline to CUDA: 19.848s

Feature levels = 2

• Inference (full pipeline): failed (AssertionError: only single scale support implemented)
• see Grouped prompt encoder for export #4 (comment)

CPU export status

• Attempted SAM3_EXPORT_FORCE_CPU=1 export for test_decoder_export_static.
• Result: failed with GuardOnDataDependentSymNode in scaled_dot_product_attention (decoder RPB cross-attn). CPU export is not currently reliable.

Model/data changes

• sam3/model/decoder.py: rework presence-token masking and cross-attn path to support RPB masks via scaled_dot_product_attention without MHA guard issues; relax coordinate cache asserts for export stability.
• sam3/model/encoder.py: fix tensor-dim check (x.dim() vs x.dim).
• sam3/model/geometry_encoders.py: guard pin_memory and fix ROIAlign input layout to avoid export issues.
• sam3/model/position_encoding.py: avoid caching with symbolic shapes to prevent export SymInt guard errors.
• sam3/model_builder.py: thread num_feature_levels into transformer/model construction for dynamic experiments (defaults unchanged).
• sam3/train/data/*: cleanup unused loop indices (no behavior change).

Notes

• Multi‑feature‑level support is limited: the decoder asserts spatial_shapes.shape[0] == 1 when box RPB is enabled, so full pipeline/decoder‑only export and inference fail for num_feature_levels > 1. This likely impacts small‑object performance relative to a true multi‑scale feature stack.
• Load timing requires torchvision.ops to be imported to register roi_align before torch.export.load.