Optim-wip: Add Neural Style Transfer tutorial & pre-trained VGG-16 model from Caffe

[ProGamerGov]

• Added the pretrained VGG-16 model from Caffe used by the Visual Geometry Group for the ImageNet ILSVRC-2014 Challenge.
• Added neural style transfer (nst) tutorial notebook for InceptionV1 and VGG models.

The InceptionV1 portion of the tutorial may fail until the input range changes from #655 are implemented. I updated the transform_input function in this PR to match the changes made in the other PR, so #655 is no longer required for this PR.

[ProGamerGov]

Some of the style images in the tutorial are currently failing to download as Wikipedia is now requiring we supply a user agent with a download request. We'll need to add a user agent header to the ImageTensor.open function to resolve the issue, or we could host the images ourselves.

[aobo-y]

Hi @ProGamerGov, learned a lot by going through the tutorial! But do you mind providing me more context? What is the goal and why do you believe Captum needs a neural style transfer tutorial? I can see it demonstrates how to leverage some newly added utility & loss functions to redo a similar PyTorch tutorial. However, why can't we show them in a tutorial with stronger taste of model interpretability?

[ProGamerGov]

Hi @aobo-y, thank you for reviewing this tutorial and it's associated code!

In terms of the benefits of including an NST tutorial, I have the following:

• Show users how to use a custom optimization function for optimizers that require closure functions. No other tutorial is able to cover this at the moment.
• Illustrate the usage and benefits of FFT image parameterization and transform robustness and show how they can conflict with architectural robustness (like in VGG models for example).
• Show users how to work with custom input ranges.
• Demonstrate the use of ImageTensor’s open() class function to load images.
• Lucid’s NST tutorial is seemingly one of it’s most popular tutorials and thus having a Captum equivalent will aid in user migrations.

[ProGamerGov]

We can probably skip this PR for now.