SAMExporter 0.5.0: SAM3 quality, more models, portable runtimes

[vietanhdev]

Summary

• update pinned official SAM, SAM2/SAM2.1, and SAM3 sources
• add EfficientSAM-Ti/S runtime support and complete MobileSAM decoder export
• fix SAM1 aspect ratios/colors, SAM2 mask quality, and SAM3 raw-head postprocessing
• use official-style SAM3 mask NMS plus prompt-focused selection and text examples
• support installed ONNX Runtime accelerator providers and provider-specific extras
• publish and document validated ONNX models at nrl-ai/samexporter-onnx-models
• retain real-model images and logs under visual_results
• add CI and prepare package version 0.5.0

Validation

• 42 unit tests pass on Python 3.13
• ruff, shell syntax, YAML, wheel, sdist, and Twine checks pass
• clean installed wheel includes the CLIP BPE resource and tokenizer attribution
• real SAM ViT-B, MobileSAM, EfficientSAM-Ti, SAM2.1 Tiny, and SAM3 ONNX runs reviewed visually
• SAM3 packaged-tokenizer output is pixel-identical to the reviewed real-model result

Issue coverage

Addresses #1, #3, #4, #5, #7, #8, #13, #15, #17, #18, #20, #21, #23, #27, #28, #30, #31, and #33. Remaining requests that need a separate design or reproducer are documented during issue triage.