Fix QNN E2E evaluation failures

[DingmaomaoBJTU]

QNN E2E started failing after transformers 5 changed CLIP export to produce per-layer SDPA attention mask guard graphs that include IsNaN and Where. Those graphs are rejected by the QNN backend during validate/finalize, so this fixes the export regression instead of disabling the affected eval targets.

Summary

• Add a reusable export context manager in transformers_compat.py that temporarily switches HF-style configs exposing _attn_implementation to eager attention during torch.onnx.export, then restores the original values afterward.
• Wrap the HTP ONNX export path with that context so transformers 5 exports avoid the QNN-incompatible SDPA guard pattern without model ID, architecture, node, or op-name hardcoding.
• Stabilize QNN E2E handling by classifying explicit QNN backend validation/finalize rejections as unsupported failures, tightening registry target filtering, and covering the related subprocess/template behavior with tests.
• Treat Windows native access violation exits from winml build as non-fatal only when the completed ONNX artifact can be resolved, so post-build teardown crashes do not hide usable artifacts.

Validation

• Targeted exporter, transformers compat, and run_eval unit pytest coverage passed.
• winml sys confirms local QNNExecutionProvider registration for NPU/GPU/CPU.
• openai/clip-vit-base-patch32 QNN GPU feature-extraction and zero-shot-image-classification E2E perf passed.
• google-bert/bert-base-multilingual-cased QNN GPU masked-lm and fill-mask E2E perf passed.
• google-bert/bert-base-multilingual-cased QNN NPU fill-mask E2E perf passed for fp16 and w8a16.
• google/flan-t5-base QNN GPU E2E perf passed.