match sdpa patterns from HF (#100609)
Adds sdpa patterns seen in HF models.
To actually make the patterns match, we need constant folding to remove addition of all-zeros mask, and figure out what to do with low mem dropout.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/100609
Approved by: https://github.com/jansel
Author
Natalia Gimelshein
