[Feature] Investigate other gene embeddings

[quentinblampey]

So far, we use the gene embeddings of scGPT for the initialization of the n_genes x embedding_size embedding matrix in novae. See model.py and embed.py.

Meanwhile, other single-cell foundation models have been created. We could see if their gene embedding (not the model itself, just the matrix of embeddings) could be better. If so, we could try replacing this initialization matrix with a better one.

NB: By "better", I mean with a higher variability between different genes. We could also make a UMAP of the gene embeddings to see if we get a better proximity between related genes, and a higher separation between highly-different gene families.