[ghstack-poisoned]

Try to fix #138772 .

aten._scaled_dot_product_efficient_attention_backward requires the out and gradient_out to have stride order (3, 1, 2, 0).  When Inductor layout optimization is enabled, Inductor may change tensor strides if they are not user visible. For efficient_attention_backward, Inductor tries to follow eager strides. But the eager strides Inductor gets for backward graph may be the one after optimization. There are a few possible fixes:
1. change the kernel to allow stride order other than  (3, 1, 2, 0). This is probably hard
2. backout https://github.com/pytorch/pytorch/pull/112045/files and don't do layout optimization if the model contains efficient_attention. 
3. Force (3, 1, 2, 0) strides order for the relevant tensors
4. Pass original eager layouts to Inductor for the backward graph. Let Inductor follow those layouts for tensors with extra layout requirement.


The PR implements option 3. Option 4 looks more general to me, I think we can do this in long term.

I tried to add a test but failed to repro: https://gist.github.com/shunting314/fe37a246aad269de9ea00199446688f6

Here is the original command to repro the issue:
```
TORCHINDUCTOR_LAYOUT_OPTIMIZATION=1 PYTORCH_NO_CUDA_MEMORY_CACHING=1 CUDA_LAUNCH_BLOCKING=1 time python benchmark.py --model maxvit_nano_rw_256 --precision bfloat16 --torchcompile --bench train --no-retry -b 64
```
benchmark.py is https://github.com/huggingface/pytorch-image-models/blob/main/benchmark.py


cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang aakhundov 

[ghstack-poisoned]