use LR finder algorithm to find optimal learning rates for different datasets and training runs

[alexshepard]

instead of doing what we do now, doing a grid search and then picking something that seems sane. from https://docs.fast.ai/learner.html#learner originally I think?