Refactor and create Graph Neural Network models

[JesusPinedaC]

This pull request aims to ensure compatibility with Torch Geometric (TG) Data objects, the most widely used data type to represent graphs in Torch. The changes are specified in the following commits:

• [878c55b]: The Graph Convolutional components were modified to be compatible with the naming convention used by TG.
• [dea3c5a]: The Message Passing components were modified to be compatible with the naming convention used by TG.
• [efa0430, 6a45885]: The blocks used in the Graph Neural Network components have been refactored to ensure compatibility with TG.
• [15947d2]: Object type checks have been included in DeeplayModule, and the read/write operations were updated to work with all supported data structures.
• [c88dd79, 6533044]: The read/write operations in Parallel have been updated to work with all supported data structures.

Moreover, this PR includes five new Graph Neural Network Models:

• Message Passing Model (MPM): A model with a "universal" nature, intended to be used as a basis to create future implementations of Graph Neural Networks, including both Graph Convolutional and Message passing.
• Graph-To-Global Message Passing Model (GraphToGlobalMPM): A model that transforms a graph into a global embedding, encoding graph-level information.
• Graph-To-Node Message Passing Model (GraphToNodeMPM): A model that transforms a graph into node embeddings, encoding node-level information.
• Graph-To-Edge Message Passing Model (GraphToEdgeMPM): A model that transforms a graph into edge embeddings, encoding edge-level information.
• Residual Message Passing Neural Network Component: Message Passing Neural Network that implements skip connections between input and output layers, enabling direct pathways for information flow.

Finally, this PR adds a new component:

• AddDict: Performs element-wise addition of values associated with user-specified keys in two dictionaries. It's designed for use in skip connections, enabling selective integration of information from different layers while maintaining the structural integrity of the input dictionaries.

All unit tests related to the changes have been updated. Additionally, I have carefully validated the implementation on the molecular property prediction examples, and it works as expected.

[BenjaminMidtvedt]

There are a lot of blocks that we'll probably need to migrate into the new structure at a later time. Other than that it all looks great!

[giovannivolpe]

@JesusPinedaC could you check the conflicts before I merge it?