poolformer_m36 correctness regression

[desertfire]

This is actually a regression from pytorch/pytorch#87054

Repro:

python benchmarks/dynamo/timm_models.py --training --accuracy --device cuda --inductor --float32 --only poolformer_m36

Error:

[2022-11-14 16:28:33,509] torch._dynamo.utils: [ERROR] RMSE (res-fp64): nan, (ref-fp64): 0.00001 and shape=torch.Size([96, 3, 7, 7])
[2022-11-14 16:28:33,509] torch._dynamo.utils: [ERROR] Accuracy failed for key name patch_embed.proj.weight.grad

[desertfire]

Assigne to this week's oncall (@eellison) to follow up.

[ngimel]

Hopefully will be fixed by #88927, but it's also flaky, even before #88927 I couldn't repro locally.

[desertfire]

Confirmed #88927 fixes this!