Debug for feature.NormalizeFeatures
Debug for autogl.datasets.utils

Debug for GeneralStaticGraph -> pyg Data conversion
Debug for autogl.datasets.utils

Revise DGL dataset to unify names for data of graph
Debug for data conversion utility
Debug for autogl.datasets.utils.graph_get_split

Revise MUTAG dataset for DGL backend
Debug for OGB nodes datasets

Debug for Default Implementation of GeneralStaticGraph
Debug for OGB nodes datasets
Improvement for feature generators and graph feature extractors to support any generic provided conventional data

Refactor of Dataset and Feature Engineer to provide a unified API for multi-backend.
In particular, this refactor introduces a unified dataset abstraction with novel feature engineer implementation.

Merge from refactor_test branch and already fixed bugs detected through tests for refactor, and temporarily override  method of dataset provided by PyTorch-Geometric to maintain compabillity with PyTorch-Geometric>=1.7.0

Temporarily override the get method for the datasets provided by PyTorch-Geometric to make the AutoGL compatible with PyTorch-Geometric>=1.6.2.
For version 1.6.2 or higher, the PyG InMemoryDataset has an additional internal attribute __data_list__ by default. If an in-memory dataset has the attribute and it is not None, then the item in the given specific index will be directly returned without accessing the data property.
Besides, the PyG matainer rusty1s also mentioned that "accessing dataset.data is not recommended" in https://github.com/rusty1s/pytorch_geometric/issues/1984.
In other words, mutating the data property of InMemoryDataset is not recommended by the PyTorch-Geometric. However, the current version of AutoGL does modify the data property of dataset.
Thus the get method of the PyG provided datasets is overrided to delete the __data_list__ attribute in advance before calling the virtual get method, aiming to temporarily fix the compatibility issue.

In addition, the ModelNet10 and ModelNet40 datasets are significantly modified.