[Train] Update documentation to show how to use `torchmetrics` for metric aggregation

[Yard1]

Description

For some applications of Ray Train, it may be desirable to report metrics from all workers and/or report aggregations, such as mean and std. While Train currently reports metrics only from rank 0 worker, Torch users can use torchmetrics to aggregate metrics from other workers and report them on rank 0. We should update the documentation to show how that can be done.

Use case

No response