Enable building an ensemble model from the cross validation checkpoints of a BYOL model

[dumbledad]

The method MLRunner.run_inference_for_lightning_models takes a list of checkpoint paths as an argument, but then makes sure that there is only one used (here):

    if len(checkpoint_paths) != 1:
        raise ValueError(f"This method expects exactly 1 checkpoint for inference, but got {len(checkpoint_paths)}")

We want to change this so that the checkpoints gleaned from a BYOL cross validation run can be used as an ensemble model.

• This will use the methods defined in the InnerEyeInference abstract class. We expect that these methods are sufficient, but an extension or redesign may be required.
• Use the simple linear regression model as the basis of an exemplar.
• It would be great to use another, more realistic model. The UHB Covid-19 model is not a lightning model, how about the Single Cell model?

AB#4219

[dumbledad]

Is this related to #377?