imagenet: add rank indicator to progress summary when distributed

[hartb]

The imagenet example supports distribution. When run in distributed mode
multiple ranks will be performing training and testing, and all will be
producing progress meter output. The output from the various ranks is
interleaved, and it's ambiguous which rank produced any specific bit of
progress output.

This change adds a rank indicator to the progress prefix when
distribution is in force:

Non-distributed:

$ python main.py --epochs 1 ...
=> creating model 'resnet18'
Epoch: [0][ 0/5005] Time 23.173 (23.173) ...
...
Test: [ 0/196] Time 10.899 (10.899) ...
...

Distributed (note the additional [0] and [1] in the prefixes):

$ python main.py --epochs 1 --dist-url 'tcp://127.0.0.1:2200'
--dist-backend 'nccl' --multiprocessing-distributed --world-size 1 --rank 0 ...
Use GPU: 1 for training
Use GPU: 0 for training
=> creating model 'resnet18'
=> creating model 'resnet18'
Epoch: [0][0][ 0/5005] Time 20.770 (20.770) ...
Epoch: [0][1][ 0/5005] Time 20.771 (20.771) ...
...
Test[0]: [ 0/391] Time 7.295 ( 7.295) ...
Test[1]: [ 0/391] Time 7.188 ( 7.188) ...
...

[hartb]

The rank indicator could be updated to include a more explicit hint of what it's for ("[R:..]").

This change omits the indicator when not distributed, it could always be included, maybe with some non-numeric value when not distributed ("[-]").

And could be a made a proper part of ProgressMeter with a bit more plumbing.