Add docs for `load_query_data` for `SCVIDataModule`

[maarten-devries]

The SCVI tutorial notebook currently only shows how to do inference on the same dataset that was used to train the model.

This is a good first step, but arguably an even more common workflow is to embed an unseen dataset using a pre-trained scVI model. I wonder if this is currently supported; if it is, then it would be great to add it to the tutorial notebook.

Traditionally, in scVI, one would run:

scvi.model.SCVI.prepare_query_anndata(adata, model_path)
model = scvi.model.SCVI.load_query_data(adata, model_path)

Currently, the SCVIDataModule errors out when it sees a value of batch_labels that it has not seen during training:

ValueError: y contains previously unseen labels: 'new_batch'

Of course, as a hacky workaround, I could just set the batch label of my new dataset to one of the seen batch labels, but this is not really biologically correct. It should be possible to specify a new unseen batch.

[aaronwolen]

Great point, @maarten-devries. We'll take a look!

[maarten-devries]

Also, please add docs for how to do inference without a SCVIDataModule dataloader. It's a great use case if you have a large dataset that doesn't fit into RAM, but what if it does fit into RAM?

[maarten-devries]

I am going to switch again to TileDBDataModule in scvi-tools, since I believe it handles all of these cases out of the box.
I think it ultimately makes sense to integrate these two or deprecate SCVIDataModule long-term.

[maarten-devries]

I spent a bit of time on this. For TileDBDataModule, we use the **setup_args loaded from the saved registry. For SCVIDataModule, this is also possible but the code will look slightly different I think.

# Load the reference model to get setup args
reference_model = scvi.model.SCVI.load(model_path, adata=False)
setup_args = reference_model.registry['setup_args']
batch_key = setup_args.get('batch_key', '_scvi_batch')
labels_key = setup_args.get('labels_key', 'label')

# Add required batch and label columns if they don't exist
if batch_key not in adata.obs.columns:
    logger.info(f"Adding batch column '{batch_key}' with placeholder value")
    adata.obs[batch_key] = 'query_batch'
if labels_key not in adata.obs.columns:
    logger.info(f"Adding label column '{labels_key}' with placeholder value") 
    adata.obs[labels_key] = 'query_label'

# Use standard scVI load_query_data workflow
scvi.model.SCVI.prepare_query_anndata(adata, model_path)
scvi.model.SCVI.setup_anndata(adata, **setup_args)
model = scvi.model.SCVI.load_query_data(adata, model_path)
model.is_trained = True