Make RunningAverage and Rouge serializable

[sadra-barikbin]

Now state_dict could serialize metric attributes that themselves are metric or simple collections of metrics.(For Rouge). If an attribute is None, state_dict sets the same in the dict (for _value in RunningAverage)

Follow-up PR to #3001

[vfdev-5]

Thanks @sadra-barikbin, PR looks good to me!
Can you please check and report here how it works on MNIST while saving and resuming.
As a test example, please use:
https://github.com/pytorch/ignite/blob/11a1fba0594f6ac0db9e8bc5f21305dc338bb7c1/examples/mnist/mnist_save_resume_engine.py
Here are few instructions on https://github.com/pytorch/ignite/blob/master/examples/mnist/README.md#training-save--resume
Finally, we can also update mnist_save_resume_engine.py code adding RunningAverage metric for batch loss to the trainer and plotting it.

[sadra-barikbin]

Here is the running loss until stop and after resume:

[sadra-barikbin]

Regarding your comment in the main PR, I introduced a new if branch in Checkpoint to accept a dictionary of serializable objects as an item in the to_save. This way we could do

metrics = {'acc': Accuracy(), 'nll': Loss()}
to_save={'model':model, 'trainer':trainer, 'metrics': metrics}

How is this?