Support metrics reduction outside of DDP

[ytang137]

Do we currently have a mechanism to allow reducing multiple metrics objects out of the context of DDP? An example of use case is doing distributed evaluation using frameworks other than DDP, where each worker has a metric object that keep track of the state for the worker's shard of data: metrics = [metric_0, metric_1, metric_2, metric_3, ...]. To compute the final metric across the entire dataset, we need a mechanism to reduce the metrics. By looking at the torchmetrics/src/torchmetrics/metric.py, I see one potential solution:

metric_reduced = Metric()
for metric in metrics:
    metric_reduced._reduce_states(metric.metric_state)

# Compute final metric
final_metric = metric_reduced.comput()

Is this a reasonable approach other than the fact that_reduced_state is meant to be private? Is there a better approach? Thanks.