Add DiceCELoss as a metric

[matt3o]

Is your feature request related to a problem? Please describe.
Since DiceCELoss is available as a loss, it would make sense for it to be available as a metric too.

Describe the solution you'd like
Either a dedicated DiceCELoss metric like e.g. MeanDice for the DiceLoss OR a pluggable DiceCELossMetric which can be used as input to IgniteMetric.

Describe alternatives you've considered
Passing the DiceCELoss into IgniteMetric results in an error since it does not have a reset button. Which kinda makes sense, this is no metric, it does something different.

Additional context
I am also curious about why the dice got implemented at two places, namely MeanDice and the DiceLoss. From a quick glance I believe they do the same computation - so is there any reason for the code to duplicated here? The calculation logic for both should be the same.

[wyli]

This function actually wraps any loss as part of cumulative metric https://github.com/Project-MONAI/MONAI/blob/dev/monai/metrics/loss_metric.py perhaps we can focus on this class.

Meandice is not differentiable and (thus) is much faster than dice loss.

[matt3o]

@wyli See #6695 about that topic, I tried to implement it with the class you mentioned. Not yet good code, but it's working on my setup at least.

[matt3o]

@wyli Just fyi: I did some experiments and actually the loss is 5 times faster for my use case. (label of size num_classes, H > 224, W > 224, D > 224, all calculations on the GPU)
With with torch.no_grad(): it is a lot faster than the metric plus it has a is a bit easier to use (Nan workaround with ignore_empty=False not needed).

[wyli]

addressed in #6695, thank you @matt3o