Accuracy incurs a GPU/CPU sync point

[ndrwrbgs]

🐛 Bug

If running Accuracy inside the training loop, training is slowed due to CPU/GPU sync point

To Reproduce

Steps to reproduce the behavior...

Pull Request includes reproduction, a unit test which fails before the modification and passes after.
#1 Update _safe_divide to allow Accuracy to run on the GPU

P.S. I opened it in the fork due to the question # 1 under "Before submitting", but Contributing.md does not mandate discussion prior "Submit a PR!", would it be possible to update one or the other with the latest guidance?

Expected behavior

Accuracy can be calculated on the GPU, without syncing

Environment

• TorchMetrics version (and how you installed TM, e.g. conda, pip, build from source): conda torchmetrics 1.4.0.post0 pyhd8ed1ab_0 conda-forge
• Python & PyTorch Version (e.g., 1.0): python 3.11.9 hb806964_0_cpython conda-forge
• Any other relevant information such as OS (e.g., Linux):

[github-actions]

Hi! thanks for your contribution!, great first issue!

[SkafteNicki]

Hi @ndrwrbgs, thanks for reporting this issue and thanks for already having a fix and unittest for the bug!
Could you open a PR directly to the main torchmetrics repository so we can get the fix incorporated?

[ndrwrbgs]

Thank you, all done at #2640