`bench`: benchmark `SheafNN` against a GNN over graph-structured dataset

[FiberedSkies]

Description

Create a comprehensive benchmark to evaluate SheafNN performance against standard Graph Neural Network (GNN) architectures on graph-structured datasets. Following the insights provided by the original Sheaf Neural Networks paper, its expected that Koho will outperform typical GNNs on heterophilic datasets or graphs with asymmetric relationships between nodes. This can most easily be benched with learnable restrictions, in which case this will likely require significant data. essentially instead of $1$ diffusion weight matrix, theres now $1+M$ matrices to learn, where $M$ is the number of total restrictions in the graph sheaf.

Details

• pick a dataset with known GNN benchmarks, and implement a preprocess using Graph builder pattern from feat: implement Graph with canonical orientations #12
• choose a hyperparameter optimization method to derive a good start point, and split the data for training and testing
• write a benchmark module using the mini-batch training from feat: implement mini-batch training support #11 over the preprocessed dataset