Visual-graph-datasets

-------------------

- Slightly changed the generation process of the "rb_adv_motifs" dataset.
- Added the class of experiments based on ``experiments.csv_sanchez_lengeling_dataset.py``, which convert
the datasets from the paper into a single CSV file, which can then be further processed into a visual graph
dataset.
- Added utility function ``util.edge_importances_from_node_importances`` to derive edge explanations from
the node explanations in cases where they are not created.
- Started to move towards the new pycomex Functional API with the experiments
- Added more documentation to ``typing``

Model Interfaces and Mixins

- Added the ``visual_graph_datasets.models`` module which will contain all the code which is relevant for
models that specifically work with visual graph datasets
- Added the ``models.PredictGraphMixin`` class, which is essentially an interface that can be implemented
by a model class to signify that it supports the ``predict_graph`` method which can be used to query a
model prediction directly based on a GraphDict object.

Examples

- Added a ``examples/README.rst``
- Added ``examples/01_explanation_pdf``

-------------------

- Changed the version dependency for numpy

-------------------

- Fixed a bug in ``utils.dynamic_import``
- Fixed a bug in ``data.load_visual_graph_element``

-------------------

- Changed the "config" cli command to also be usable without actually opening the editor. This can be used
to silently create or overwrite a config file for example.

-------------------

Implemented the "preprocessing" feature. Currently a big problem with the visual graph datasets in general
is that they are essentially limited to the elements which they already contain. There is no easy way to
generate more input elements in the same general graph representation / format as the elements already in
a dataset. This is a problem if any model trained based on a VGD is supposed to be actually used on new
unseen data: It will be difficult to process a new molecule for example into the appropriate input tensor
format required to query the model.

The "preprocessing" feature addresses this problem. During the creation of each VGD a python module
"process.py" is automatically created from a template and saved into the VGD folder as well. It contains all
the necessary code needed to transform a domain specific implementation (such as a SMILES code for example)
into a new input element of that dataset, including the graph representation as well as the visualization.
This module can either be imported to use the functionality directly in python code. It also acts as a
command line application.

- Added the base class ``processing.base.ProcessingBase``. This class encapsulates the previously described
pre-processing functionality. Classes inheriting from this automatically act as a command line interface
as well.
- Code for a standalone python module with the same processing functionality can be generated from an
instance using the ``processing.base.create_processing_module`` function.
- Added the class ``processing.molecules.MoleculeProcessing``. This class provides a standard implementation
for processing molecular graphs given as SMILES strings.
- Added unittests for base processing functionality
- Added unittests for molecule processing functionality

- Extended the function ``typing.assert_graph_dict`` to do some more in-depth checks for a valid graph dict
- Added module ``generation.color``. It implements utility functions which are needed specifically for the
generation of color graph datasets.
- Added the experiment ``experiment.generate_rb_adv_motifs`` which generates the synthetic
"red-blue adversarial motifs" classification dataset of color graphs.

------------------

- changed dependencies to fit together with ``graph_attention_student``
- Added experiment to generate ``aggregators_binary`` dataset