We want to push changes to build Graphium 3.0, which will enable faster and more memory-efficient training and inference while also removing some of the codebase's "uglier" parts and version constraints.

• Current constraints about cuda-version=11.2 make the package not really usable.
• The point above is due to torchmetrics >=0.7.0,<0.11 constraint. That constraint needs to be relaxed, which requires to remove the file ipu_metrics.py and to change from functional metrics to class metrics.
• Moving to a C++ molecular featurization for super-fast at-dataloading featurization of molecules. The caching will be optimized and only contain the labels.
• Support for multi-gpu, and making sure the metrics and loss sync correctly across devices.
• Standardizing pre-nn and pre-nn-edges to be part of the MLPEncoder and EncoderManager
• Fix the issue with multiple node ordering coming from multiple tasks that require different orders (nodes, edges, etc.)