Enable multi-gpu AnomalyScoreThreshold

[Seanny123]

Description

compute is not supposed to update state variables, in this case value, as defined by self.add_state("value", default=torch.tensor(default_value), persistent=True). The current implementation doesn't cause problems until you use multiple GPUs, where any changes to self.value in compute are discarded. Updating the variable is supposed to happen in update() or outside of the class. I elected to do it outside of the class.

• Fixes [Bug]: AnomalyScoreThreshold is incompatible with multi-GPU training #1398

Given how awkward this was to figure out and the hacky fix required, maybe AnomalyScoreThreshold shouldn't sub-class what its currently sub-classing?

Checklist

Ensure that you followed the following

• I have performed a self-review of my code

[samet-akcay]

@ashwinvaidya17 and @djdameln, you guys are expert at this. Can you check this out?

[Seanny123]

As discussed in the linked issue, this seems to hurt model accuracy for some reason? So maybe don't merge.