Add UMAP embedding visualisation script

[amorehead]

• Adds a visualise.py script that lets one load any pre-trained encoder checkpoint, embed an entire dataset as a collection of graph embeddings, and then plot the collected embeddings using a Gaussian flavour of UMAP.
• Lists the top 20 most common label string values in the legend of the resulting figure. For example, when running this script with the settings python proteinworkshop/visualise.py ckpt_path=$MY_PT_CA_BB_CKPT_PATH plot_filepath=output_visualisations/pt_encoder_fold_superfamily.png dataset=fold_superfamily encoder=gcpnet features=ca_bb task=multiclass_graph_classification, this yields the top 20 most common superfamilies being identified in the figure's legend.
• Plots may be adjusted as desired by the user. The core logic for acquiring dataset graph embeddings is provided (i.e., batteries included).

[amorehead]

I am currently facing an odd indexing error when I run for the full fold_superfamily dataset. On a subset of the dataset, I am seeing clusterings such as the following, for reference.

[amorehead]

A note for later: One idea for quantifying each encoder's clustering ability would be to report each method's Dunn Index and Silhouette value.

[a-r-j]

LGTM so far. I'll add a property to the datamodules for mapping labels to their corresponding text where possible.

[a-r-j]

How are devices handled? Normally the L.Trainer would handle this but we're not using that here. I suppose this is all happening on the CPU?

[amorehead]

Great point. I completely missed that the trainer wasn't being instantiated here. No wonder my initial runs of the script were so slow. In this commit, I've added GPU support for both embed.py and visualise.py. Moreover, after timing visualise.py with GCPNet on the fold_superfamily dataset, I was able to create the plot below in around 10 minutes using a GPU.

Overall, it looks like the b.1 superfamily (i.e., the most common superfamily in the dataset) is clustered pretty well.