Add support for PolyLoss in torchvision

[frgfm]

🚀 The feature

The goal would be to add a PolyLoss PyTorch module behaving exactly the same way as any PyTorch core loss module. And the implementation would follow the following paper: https://arxiv.org/abs/2204.12511

One question still pending though: where should this be located in torchvision?

Motivation, pitch

As discussed in #6323, some losses have popped up over the last year. Before they can make their way to core, we should integrate those in torchvision first.

Alternatives

No response

Additional context

No response