Add ReplaceAttentionMaskValue graph surgeon, Bert QNN EP example

[jambayk]

Describe your changes

• New graph surgeon that replaces the default attention mask value (transformers uses min float value) with a different value. The min float value does not quantize well and leads to -inf when dequantized. We use a value like -10000 to make it easier to quantize.
• Bert example for qnn ep added. Accuracy is comparable to float model.
  • Latency on cpu: ~900ms
  • Latency on npu (Snapdragon XElite): ~8ms
     

Checklist before requesting a review

• Add unit tests for this change.
• Make sure all tests can pass.
• Update documents if necessary.
• Lint and apply fixes to your code by running lintrunner -a
• Is this a user-facing change? If yes, give a description of this change to be included in the release notes.
• Is this PR including examples changes? If yes, please remember to update example documentation in a follow-up PR.

(Optional) Issue link