Summary:
This changes ClassificationTask to compute some high-level performance
numbers (img/sec) and plot them in Tensorboard. This is useful for comparing
performance optimizations since we now get a "blessed" performance number.

Also, this was done in a way that's comparable to NVidia's benchmarks (e.g. https://ngc.nvidia.com/catalog/model-scripts/nvidia:resnet_50_v1_5_for_pytorch/performance), so we know how well we're doing compared to other implementations.

In terms of implementation, I could have made a hook instead, but decided against it for two reasons: (1) it would introduce dependencies between hooks; (2) we want to control precisely when the timing measurements are taken;

Pull Request resolved: facebookresearch#385

Test Plan: ./classy_train.py --config configs/template_config.json

Reviewed By: mannatsingh

Differential Revision: D19739656

Pulled By: vreis

fbshipit-source-id: a63c394308851e6accee9d260d9cb1d972f33a7f