[PyTorch] MHA: fix contiguity assumption in transform_bias_rescale_qkv (#72465)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/72465
This code path incorrectly assumed input tensors were contiguous. Now we check that.
ghstack-source-id: 149201476
Test Plan: CI
Reviewed By: ngimel
Differential Revision: D34007665
fbshipit-source-id: c43438f2495e32304ea3f7846e01eceb4a9448f7
(cherry picked from commit 0767b225f23846c1636ac3622f46b0c5ec071d96)
