deprecate input_batch in model inputs

[tianyu-l]

Now that we are creating FlexAttention block masks outside the model, the input_batch field in model inputs is no longer needed.

This PR also renames extra_args to be extra_kwargs, as it's a dictionary of kwargs, technically.