Make ocrvqa default in the_cauldron_dataset

[RdoubleA]

So that we can remove the subset argument from the configs and it's much easier to override for a custom dataset via CLI, i.e., multimodal_chat_dataset. Otherwise you will need to use the CLI removal which is awkward.

[joecummings]

It should be as easy as the following code snipped in the cauldron dataset:

def the_cauldron_dataset(
    model_transform: Transform,
    *,
    subset: str = "orcvqa",
	...
)