Landan/deeplabv3 ade20k example

[Landanjs]

What does this PR do?

Contributes example script to train DeepLabv3+ on ADE20k.

Old WandB runs for the full 128 epochs and without any constant LR:

	mIoU	Time-to-Train (hours)
Baseline + dice	46.3	3.85
Mild	46.84	3.31
Medium	49.13	5.0
Hot	48.22	6.68

I plan to rerun with constant LR and the target number of epochs for each recipe

[Landanjs]

Updated description with previous WandB runs, will attempt new runs with constant LR and appropriate recipe number of epochs

[coryMosaicML]

Great work! A few minor comments.

Biggest things:
Maybe include a little info on how to install things, or which docker image to use
One of the params for the hot recipe was a bit off.