Add support for Python 2

[fmassa]

This PR adds support for Python 2, and should fix #7

I've successfully run tools/train.py and tools/test.py after these fixes, and training / inference succeeded without errors.

I'll run a quick schedule train/test cycle to verify that the numbers match the expected results.

[fmassa]

Thanks for the review @rbgirshick !

I'm performing some full trainings to verify that accuracy is the same.
Training on the reduced scheduled gave expected results, as well as inference on pre-trained models.
Once I verify that this works as expected on the full training I'll merge it.

[fmassa]

Here are the results for training on Mask R-CNN R-50 FPN 1x. Accuracies match the baselines, and it seems that the Python 2 runtime is a bit slower. I've run it twice.

backbone	type	lr sched	im / gpu	train mem(GB)	train time (s/iter)	total train time(hr)	inference time(s/im)	box AP	mask AP	model id
R-50-FPN	Mask	1x	2	5.2	0.4657	11.6	0.13757 + 0.034	37.7	34.3	6486235
R-50-FPN	Mask	1x	2	5.2	0.4670	11.6	0.14075 + 0.034	37.6	34.2	6486657

I'm merging this.