Fix Transformers 4.57 export compatibility

[Yue Sun (KayMKM)]

Summary

• normalize integral attention masks before Optimum calls PyTorch SDPA
• use the model's positional export protocol during HTP hierarchy tracing
• pass a Transformers 4.57-compatible 3-D binary mask to the BLIP decoder

Validation

• 54 focused unit tests passed
• DistilBERT fill-mask evaluation passed on OpenVINO NPU
• BLIP FP16 image-to-text evaluation passed on OpenVINO NPU
• both affected eval E2E cases passed in 445.90s
• Ruff passed