Add SDPA pattern for HuggingFace models BF16 (#121202)
### Description
- Add pattern for bf16 input type with fp32 attention mask. (Example model: ElectraForCausalLM)
- Add pattern with batch_size=1 to avoid some clones in graph. (Example model: text-classification+prajjwal1-bert-tiny)
### Newly matched models
Dtype: bf16, machine: SPR
#### Dynamo HuggingFace models
- ElectraForCausalLM (speedup=2.09x)
- ElectraForQuestionAnswering (speedup=4.22x)
- AlbertForQuestionAnswering (speedup=1.36x)
- AlbertForMaskedLM (speedup=1.39x)
#### OOB HuggingFace models
- multiple-choice+google-electra-base-discriminator
- text-classification+prajjwal1-bert-tiny
- text-classification+prajjwal1-bert-mini
- text-classification+google-electra-base-generator
- text-classification+bert-large-cased
- casual-language-modeling+xlm-roberta-base
- text-classification+roberta-base
- text-classification+xlm-roberta-base
- text-classification+albert-base-v2
- token-classification+google-electra-base-generator
- masked-language-modeling+bert-base-cased
Pull Request resolved: https://github.com/pytorch/pytorch/pull/121202
Approved by: https://github.com/jgong5, https://github.com/jansel