Implement Gradient Shapley

Algorithm 2 of _Ghorbani, Amirata, and James Zou. [Data Shapley: Equitable Valuation of Data for Machine Learning](http://proceedings.mlr.press/v97/ghorbani19c.html). In International Conference on Machine Learning, 2242–51. PMLR, 2019._

Note that this is not a true approximation to Shapley value, since it reuses the computation of the utility on a subset to compute it on another. In particular, the marginal utility is the difference in performance after one SGD step, which is not equivalent to a full retraining. As such this breaks several assumptions, e.g. that the sequence in which one adds samples to a coalition does not affect its utility, or that the utility computations are independent.