Keras-vis

- Improved performance of guided backprop with subsequent calls.
- various minor bug/doc fixes.

- Introducing grad_modifiers, backprop_modifiers, input_modifiers.
- Implemented guided and rectified backprop.
- API changes. Deprecated visualize_regression variant and removed RegressionTarget loss,
- Added visualize_xxx_with_losses variant for flexibility.
- Moved all examples to notebooks.
- Introduced backend abstraction for backend specific code.
- Modularized visualization code.

- Implemented regression variant for all visualizations.
- Refactored common code for various visualizations for advanced use cases of specifying custom loss.
- Added RegressionTarget experimental loss function.
- Added an example use-case for visualizing a toy self-driving car.
- Added the ability to specify input range for activation visualizations
- Decoupled heatmap overlay in attention visualizations as not all inputs are images.

- Added support for N-dim attention, and activation maximization
- Huge perf improvements
- Removed a ton of unwanted dependencies.