[WIP] Explain PyTorch neural nets with Grad-CAM

[teabolt]

This PR explains image and text classifiers built in PyTorch using the Grad-CAM method, building on #315 and #325.

Images example:
Using the pretrained mobilenet_v2 network from torchvision and calling eli5.show_prediction(model, doc, image=img)

We get the classical explanation for 'dog':

Text example:
Using an example model from https://www.kaggle.com/ziliwang/pytorch-text-cnn for an insincere question classification task (https://www.kaggle.com/c/quora-insincere-questions-classification/overview), we can write eli5.show_prediction(model, doc, tokens=tokens, layer=layer, relu=False).

To get an explanation like this (green = 'insincere', red = 'neutral'):

This PR only provides basic PyTorch support.

TODO items:

• Consistency with Keras support and features.
• Image tutorial.
• Text tutorial (need to scale model down).
• Image integration tests.
• Text integration tests.
• Unit tests.
• Docstrings and docs.
• CI and coverage.
• Reviews.

[codecov-commenter]

Codecov Report

Merging #327 into master will decrease coverage by 3.20%.
The diff coverage is 71.95%.

@@            Coverage Diff             @@
##           master     #327      +/-   ##
==========================================
- Coverage   97.32%   94.12%   -3.21%     
==========================================
  Files          49       56       +7     
  Lines        3142     3472     +330     
  Branches      585      645      +60     
==========================================
+ Hits         3058     3268     +210     
- Misses         44      162     +118     
- Partials       40       42       +2     

Impacted Files	Coverage Δ
eli5/keras/__init__.py	100.00% <ø> (ø)
eli5/pytorch/gradcam.py	24.13% <24.13%> (ø)
eli5/pytorch/explain_prediction.py	29.16% <29.16%> (ø)
eli5/pytorch/image.py	30.00% <30.00%> (ø)
eli5/formatters/image.py	98.75% <92.85%> (-1.25%)	⬇️
eli5/nn/gradcam.py	95.83% <95.83%> (ø)
eli5/__init__.py	84.61% <100.00%> (+1.28%)	⬆️
eli5/base.py	100.00% <100.00%> (ø)
eli5/keras/explain_prediction.py	98.70% <100.00%> (+2.31%)	⬆️
eli5/keras/gradcam.py	100.00% <100.00%> (ø)
... and 18 more