Fix small issue with train/testing examples

[mkolodner-sc]

Scope of work done

This PR addresses two issues:

• If should_skip_training is True, we try to log a train_start_time after testing, but it has not been initialized, which causes error.
• If should_skip_training is True, we still try to re-save the model to the model uri, which should only be done after training.

We fix this by moving the train_start_time and model saving code to the code which runs if should_skip_training=False, and we introduce a time for logging how long the testing takes.

There is also a small consistency change here where we have the code block:

    if torch.cuda.is_available():
        torch.cuda.empty_cache()  # Releases all unoccupied cached memory currently held by the caching allocator on the CUDA-enabled GPU
        torch.cuda.synchronize()  # Ensures all CUDA operations have finished
    torch.distributed.barrier()  # Waits for all processes to reach the current point

which is currently done before calling .shutdown() in train and done after calling .shutdown() in test. It's good to be consistent between the two here, and this block should be called before .shutdown() in both cases so that we are only shutting down the dataloaders once it is safe to do so (all processes have reached the shutdown stage) and all cached memory is cleaned up before references are lost in .shutdown() call.

Where is the documentation for this feature?: N/A

Did you add automated tests or write a test plan?

Updated Changelog.md? NO

Ready for code review?: NO