Lightning CLI is incompatible with models defined by data

[ml-mountainman]

Is there an easy way to make a Lightning CLI work with a Lightning Module defined by data? This seems like a very common design pattern.

For example (from the docs) it doesn't appear possible to easily convert the following to a Lightning CLI:

# init dm AND call the processing manually
dm = ImagenetDataModule()
dm.prepare_data()
dm.setup()

model = LitModel(out_features=dm.num_classes, img_width=dm.img_width, img_height=dm.img_height)
trainer.fit(model, dm)

The current CLI implementation works for model re-loading. However, it requires data-dependent attributes to be specified to the config files prior to fitting, which does not seem ideal.

[ml-mountainman]

Resolved: see #9473