fix(optrace): include CPU fallback operators in QNN results

[xieofxie]

Summary

• enable ONNX Runtime profiling for QNN op-trace sessions and merge CPU fallback node timings into the final operator result
• preserve optimized node identity, exclude warmup samples, recalculate mixed-provider percentages, and label CPU operators
• persist and support offline parsing of the ORT profile artifact

Validation

• 260 affected unit tests passed
• DenseNet QNN NPU trace completed with 27 QNN operators and 15 CPU fallback operators

Fixes #1362