A better way to handle DDP raised error

With the BundleWorkflow interface, we modified some bundles to use this new feature. For example:
https://github.com/Project-MONAI/model-zoo/blob/dev/models/spleen_ct_segmentation/configs/multi_gpu_train.json

However, follow up the discussion in: https://github.com/Project-MONAI/model-zoo/pull/369 users may meet DDP error when calling the initialize or finalize if torch.distributed is not available. According to current settings, users may meet errors like:

AttributeError: module 'torch.distributed' has no attribute 'is_initialized'

The error message is not informative enough, and we may need a better way to handle this situation.

Thanks @wyli very much for posting this case.

Hi @Nic-Ma @holgerroth , do you have any suggestions on it? So far we can still keep the current settings, but it may be a place that can be enhanced in the next version.
I discussed with @wyli and maybe we can add some utility functions to handle DDP errors in the initialize and finalize steps. What do you think?