Bugfixes to Pytorch implementation

[rfriedman22]

The implementation for PyTorchDeepExplainer no longer works. I updated the code to reflect what is in the main shap repository to bring functionality up to speed since the repo was forked.

• Fixed the order that forward and backward hooks are registered.
• Updated calculation of the gradient with torch.autograd.grad to account for proper indexing.
• No longer delete attributes x and y from modules that hook to maxpool or nonlinear_1d.
• Fix calculation of gradient input size for MaxPool1d.

I also added support for the Exp function. Pytorch does not contain a Module for this function, but I use a custom implementation of this Module in my models.