[optim] Adam defaults to fused when CUDA + differentiable=False (#90865)
Step 1 in faster default optimizers.
Preliminary benchmarks show gaps in improvement on CUDA for BERT_pytorch and resnet18:
![image](https://user-images.githubusercontent.com/31798555/207707118-14221802-77ce-4ee0-96e3-04638c07924c.png)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/90865
Approved by: https://github.com/albanD