Generalize `TestTimeAugmentation` to non-spatial predictions

[bcdarwin]

Currently monai.data.TestTimeAugmentation seemingly assumes the inverse transformations can be applied to the predictions, which necessarily means the predictions must be spatial.

Seems fairly straightforward - would be willing to write up a PR if there's interest and no reason this doesn't make sense.

[TheHuntsman4]

Hello, I don't see any activity on this issue. Would be happy to take this up.

[TheHuntsman4]

@KumoLiu I had a question regarding this issue. As I have understood this, if the output is non-spatial, ie something like a scalar value like a classification score, there is no need to go about applying the inversion step.

The current solution that I had thought about for this was to have a flag parameter "apply_inverse_to_pred" that would be passed to the class that would dictate whether the inversion step has to be done or not.