Is your feature request related to a problem? Please describe.
Some metrics need to be computed over the whole dataset. To name a few metrics:

Classification: ROC/AUC
Segmentation: Dice (when training on patches the dice can not be computed for background patches -> need to compute intermediate values and save them)
Detection: mAP, FROC

For most of these examples it is not feasible to store the target and the prediction over many iterations (e.g. storing all the ground truth and predicted segmentations), so the computation needs to be splitted in multiple parts. E.g. storing the tp, fp, ... per sample for the later computation (similar things can be done for the detection metrics).

Describe the solution you'd like
Create a functional interface and a class interface for the metrics (probably related to #497).

class BaseMetric:
    ...

    def __call__(self, ...):
        # call functional interface to compute metric

class PartialMetric(BaseMetric):
    ...

    def add_sample(self, ...):
        # compute intermediate results to store for later computation

    def compute(self, ....):
         # could implement something additional based on the intermediate values
        # or directly compute the target metric

Describe alternatives you've considered
//

Additional context
This Issue is probably related to/to be tackled after #497 .

After discussing the details I would like to implement this feature (also including some of the metrics mentioned above, which we could discuss separately ) :)