scVI hyperparameter tuning

Our regular procedure for generating Census scVI embeddings trains the scVI model with an up-to-date HVG h5ad, but doesn't revisit the [hyperparameters](https://github.com/chanzuckerberg/cellxgene-census/blob/main/tools/models/scvi/scvi-config.yaml) set a couple of years ago, since which Census has grown a lot (and scVI itself has likely evolved). This notebook is a record of the "autotune" done in 2023: https://github.com/YosefLab/census-scvi/blob/main/experiments/autotune/notebooks/2023_09_autotune_report.ipynb To repeat, we'd still need to recover: - detailed recipe for generating the results analyzed there (installation, commands, etc.) - expertise/judgment to choose the best parameters from the results