Respect torch.use_deterministic_algorithms

[ezyang]

🐛 Describe the bug

Some of torchvision's backward CUDA operators (e.g., roi_align_backward) use gpu atomic add, which makes the nondeterministic. It would be great if they were coded to respect use_deterministic. The easy thing to do is just make them warn, but if we can implement a Python version of the function we can also do the same playbook as in pytorch/pytorch#101115

cc @kurtamohler

Versions

master

[NicolasHug]

Could be a way of addressing #7457

The non-determinism has been around forever - @ezyang is there a time-critical need to fix it now?

[ezyang]

I'll submit the patch. I do kind of need it.