Update saliency methods to use a framework-agnostic "core" subpackage, and moves all TF1 functionality to an optional "tf1" subpackage.

[bwedin]

Changes in this PR:

• Creates two subpackages, "core" and "tf1"
• "core" is a framework-agnostic package that re-implements all existing methods (with the exception of GuidedBackprop). Each method calls a model through a user-supplied "call_model_function", which receives the inputs and any auxiliary model arguments, and returns relevant outputs (e.g. gradients for a single output neuron with respect to the input) that are used to calculate feature attributions.
• "tf1" is functionally equivalent to the prior version, so any code that uses saliency can simply use saliency.tf1 instead.
• To avoid unnecessary dependencies, tensorflow is no longer required as part of install, but is instead listed as an extra in saliency[full] and saliency[tf1].
• Adds notebooks Examples_core.ipynb and Examples_tf1.ipynb, demonstrating how to use these subpackages with a model trained on ImageNet. Examples_tf1 uses a TF1 TFHub model, while Examples_core uses a VGG16 model from tf.keras.applications.
• Gradient path methods (IG, BlurIG, XRAI) now have a batch_size parameter, which calls the model with multiple steps in a single batch. Default batch_size is 1, which mirrors previous behavior (i.e. each step is a separate model call).
• Adds new tests for methods that previously had none (Occlusion, Gradients), as well as additional tests for error handling.

Bugfixes:

• Fixed an off-by-one error in Occlusion
• Added "scikit-image" as a package dependency

Tests for this change:

• Running all tests on a local virtual environment
• Running example notebooks and verifying the resulting masks match previous notebook outputs.

[gkapish]

Nice work!

[tolga-b]

Thank you Ben, it looks great! I left some minor comments.
Could you also please update the PR title and description to list major updates that are happening. The title could be about making the library framework agnostic, and you could list the changes in the features or bugs this fixes the description as well as how you tested (e.g. ran notebooks visually compared and ran method tests). Thank you

[tolga-b]

Looks good, could you update the PR description and title please?

[tolga-b]

This also resolves #16. We may want to think about adding more notebook examples for other frameworks in future.

[tolga-b]

Please add to the description that you no longer use tf.lite models and use tf hub for tf1 and keras for the core notebooks, respectively.

[gkapish]

LGTM